diff --git a/CMakeLists.txt b/CMakeLists.txt
index c48f41b15..5e261d183 100644
--- 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: ${WASI}")
 endif()
diff --git a/cmake/wasm-component-ld.cmake b/cmake/wasm-component-ld.cmake
index 2e7976602..d0b8b417e 100644
--- a/cmake/wasm-component-ld.cmake
+++ b/cmake/wasm-component-ld.cmake
@@ -10,7 +10,7 @@ if (NOT WASM_COMPONENT_LD_EXECUTABLE)
   ba_download(
     wasm-component-ld
     "https://github.com/bytecodealliance/wasm-component-ld"
-    "v0.5.19"
+    "v0.5.21"
   )
   ExternalProject_Get_Property(wasm-component-ld SOURCE_DIR)
   set(WASM_COMPONENT_LD_EXECUTABLE "${SOURCE_DIR}/wasm-component-ld")
diff --git a/expected/wasm32-wasip1-threads/defined-symbols.txt b/expected/wasm32-wasip1-threads/defined-symbols.txt
index 4fb5d53fc..1d2f3c63c 100644
--- 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
index 34149271a..b75189dbd 100644
--- 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
@@ -285,8 +296,13 @@ __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
@@ -294,6 +310,8 @@ __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
index 8e2f62089..f4176082a 100644
--- 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
index d785794c4..2d644ca97 100644
--- 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
index fc13322b6..ceb0c9d43 100644
--- 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
index 2cca21efc..8da5d9991 100644
--- 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
index 8b0b0b746..1c7276645 100644
--- 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
index cf91ac016..077c12318 100644
--- 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
index 69f758eb6..4777ba5d6 100644
--- 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
index 7f0d92931..c37507749 100644
--- 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
index 429d0b033..8cfffc426 100644
--- 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
index 155292d0e..f13d540a0 100644
--- 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
index 88f5dcf8f..b792f8d0e 100644
--- 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;
diff --git a/libc-top-half/musl/src/internal/lock.h b/libc-top-half/musl/src/internal/lock.h
index 29787fb1f..feacd7ce8 100644
--- a/libc-top-half/musl/src/internal/lock.h
+++ b/libc-top-half/musl/src/internal/lock.h
@@ -1,15 +1,56 @@
 #ifndef LOCK_H
 #define LOCK_H
 
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-hidden void __lock(volatile int *);
-hidden void __unlock(volatile int *);
-#define LOCK(x) __lock(x)
-#define UNLOCK(x) __unlock(x)
+#include <wasi/version.h>
+
+// Defines internal locking primitives. The main elements defined here are:
+// - __lock_t: the type of a lock variable
+// - __lock_ptr_t: the type of a pointer to a lock variable
+// - __LOCK_INIT: initializer for a lock variable
+// - LOCK/UNLOCK: functions to acquire/release a lock for a critical section that contains 
+//                potential context switch points, and thus must be executed even in cooperative threading mode.
+// - WEAK_LOCK/WEAK_UNLOCK: functions to acquire/release a lock for a critical section that does not contain any context switch points, 
+//                            and thus can be no-ops in cooperative threading mode.
+
+
+#ifdef __wasi_cooperative_threads__
+    struct __waitlist_node;
+    struct __coop_lock {
+    	int owner;  // tid of owning thread, or 0 if unlocked
+    	struct __waitlist_node *waiters;
+    };
+    
+    typedef struct __coop_lock __lock_t;
+    typedef struct __coop_lock *__lock_ptr_t;
+    #define __COOP_LOCK_INIT {0, NULL}
+    #define __LOCK_INIT __COOP_LOCK_INIT
+    
+    hidden void __lock(struct __coop_lock *lock);
+    hidden void __unlock(struct __coop_lock *lock);
+
+    #define LOCK(x) __lock(x)
+    #define UNLOCK(x) __unlock(x)
+    #define WEAK_LOCK(x) ((void)0)
+    #define WEAK_UNLOCK(x) ((void)0)
+#elif defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+    typedef volatile int __lock_t;
+    typedef volatile int *__lock_ptr_t;
+
+    #define __LOCK_INIT {0}
+    
+    hidden void __lock(volatile int *);
+    hidden void __unlock(volatile int *);
+    
+    #define LOCK(x) __lock(x)
+    #define UNLOCK(x) __unlock(x)
+    #define WEAK_LOCK(x) __lock(x)
+    #define WEAK_UNLOCK(x) __unlock(x)
 #else
-// No locking needed.
-#define LOCK(x) ((void)0)
-#define UNLOCK(x) ((void)0)
+    // No locking needed.
+    #define LOCK(x) ((void)0)
+    #define UNLOCK(x) ((void)0)
+    #define WEAK_LOCK(x) ((void)0)
+    #define WEAK_UNLOCK(x) ((void)0)
 #endif
 
-#endif
+#endif
\ No newline at end of file
diff --git a/libc-top-half/musl/src/internal/pthread_impl.h b/libc-top-half/musl/src/internal/pthread_impl.h
index 0106ac385..7a056d69d 100644
--- a/libc-top-half/musl/src/internal/pthread_impl.h
+++ b/libc-top-half/musl/src/internal/pthread_impl.h
@@ -18,6 +18,8 @@
 #include "futex.h"
 
 #include "pthread_arch.h"
+#include "lock.h"
+#include <wasi/version.h>
 
 #define pthread __pthread
 
@@ -172,35 +174,24 @@ hidden int __libc_sigaction(int, const struct sigaction *, struct sigaction *);
 #endif
 hidden void __unmapself(void *, size_t);
 
+#ifdef __wasi_cooperative_threads__
+struct __waitlist_node {
+    uint32_t tid;
+    struct __waitlist_node *next;
+};
+hidden void __waitlist_wait_on(struct __waitlist_node **list);
+hidden void __waitlist_wake_one(struct __waitlist_node **list);
+hidden void __waitlist_wake_all(struct __waitlist_node **list);
+#else
 #ifndef __wasilibc_unmodified_upstream
 hidden int __wasilibc_futex_wait(volatile void *, int, int, int64_t);
 #endif
 hidden int __timedwait(volatile int *, int, clockid_t, const struct timespec *, int);
 hidden int __timedwait_cp(volatile int *, int, clockid_t, const struct timespec *, int);
 hidden void __wait(volatile int *, volatile int *, int, int);
-static inline void __wake(volatile void *addr, int cnt, int priv)
-{
-	if (priv) priv = FUTEX_PRIVATE;
-	if (cnt<0) cnt = INT_MAX;
-#ifdef __wasilibc_unmodified_upstream
-	__syscall(SYS_futex, addr, FUTEX_WAKE|priv, cnt) != -ENOSYS ||
-	__syscall(SYS_futex, addr, FUTEX_WAKE, cnt);
-#else
-#ifdef _REENTRANT
-	__builtin_wasm_memory_atomic_notify((int*)addr, cnt);
-#endif
-#endif
-}
-static inline void __futexwait(volatile void *addr, int val, int priv)
-{
-#ifdef __wasilibc_unmodified_upstream
-	if (priv) priv = FUTEX_PRIVATE;
-	__syscall(SYS_futex, addr, FUTEX_WAIT|priv, val, 0) != -ENOSYS ||
-	__syscall(SYS_futex, addr, FUTEX_WAIT, val, 0);
-#else
-	__wait(addr, NULL, val, priv);
+hidden void __wake(volatile void *addr, int cnt, int priv);
+hidden void __futexwait(volatile void *addr, int val, int priv);
 #endif
-}
 
 hidden void __acquire_ptc(void);
 hidden void __release_ptc(void);
diff --git a/libc-top-half/musl/src/internal/stdio_impl.h b/libc-top-half/musl/src/internal/stdio_impl.h
index 7f19dd951..cfd476715 100644
--- a/libc-top-half/musl/src/internal/stdio_impl.h
+++ b/libc-top-half/musl/src/internal/stdio_impl.h
@@ -8,10 +8,21 @@
 
 #define UNGET 8
 
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+#include "lock.h"
+#include <wasi/version.h>
+
+#ifdef __wasi_cooperative_threads__
+#define FFINALLOCK(f) __lockfile((f))
+#define FLOCK(f) int __need_unlock = __lockfile((f))
+#define FUNLOCK(f) do { if (__need_unlock) __unlockfile((f)); } while (0)
+#define __STDIO_LOCK_INIT {0, 0}
+#define __STDIO_LOCK_RESET(lock) do { (lock)->owner = 0; (lock)->waiters = NULL; } while (0)
+#elif defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 #define FFINALLOCK(f) ((f)->lock>=0 ? __lockfile((f)) : 0)
 #define FLOCK(f) int __need_unlock = ((f)->lock>=0 ? __lockfile((f)) : 0)
 #define FUNLOCK(f) do { if (__need_unlock) __unlockfile((f)); } while (0)
+#define __STDIO_LOCK_INIT -1
+#define __STDIO_LOCK_RESET(lock) do { (*(lock)) = -1; } while (0)
 #else
 // No locking needed.
 #define FFINALLOCK(f) ((void)(f))
@@ -51,7 +62,7 @@ struct _IO_FILE {
 #endif
 	int mode;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	volatile int lock;
+	__lock_t lock;
 #endif
 	int lbf;
 	void *cookie;
@@ -73,8 +84,9 @@ extern hidden FILE *volatile __stdout_used;
 extern hidden FILE *volatile __stderr_used;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-hidden int __lockfile(FILE *);
-hidden void __unlockfile(FILE *);
+// Needs to be unhidden because they're used by the emulated signal implementation
+int __lockfile(FILE *);
+void __unlockfile(FILE *);
 #endif
 
 hidden size_t __stdio_read(FILE *, unsigned char *, size_t);
diff --git a/libc-top-half/musl/src/locale/locale_map.c b/libc-top-half/musl/src/locale/locale_map.c
index ebfb158f4..ca25ac827 100644
--- a/libc-top-half/musl/src/locale/locale_map.c
+++ b/libc-top-half/musl/src/locale/locale_map.c
@@ -31,8 +31,8 @@ static const char envvars[][12] = {
 };
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-volatile int __locale_lock[1];
-volatile int *const __locale_lockptr = __locale_lock;
+__lock_t __locale_lock[1];
+__lock_t *const __locale_lockptr = __locale_lock;
 #endif
 
 const struct __locale_map *__get_locale(int cat, const char *val)
diff --git a/libc-top-half/musl/src/malloc/oldmalloc/malloc.c b/libc-top-half/musl/src/malloc/oldmalloc/malloc.c
index 25d00d44d..f4caed87c 100644
--- a/libc-top-half/musl/src/malloc/oldmalloc/malloc.c
+++ b/libc-top-half/musl/src/malloc/oldmalloc/malloc.c
@@ -27,6 +27,16 @@ static struct {
 
 /* Synchronization tools */
 
+#ifdef __wasi_cooperative_threads__
+/* The implementation has no yield points, so locks can be no-ops. */
+static inline void lock(volatile int *lk)
+{
+}
+
+static inline void unlock(volatile int *lk)
+{
+}
+#else
 static inline void lock(volatile int *lk)
 {
 	int need_locks = libc.need_locks;
@@ -43,6 +53,7 @@ static inline void unlock(volatile int *lk)
 		if (lk[1]) __wake(lk, 1, 1);
 	}
 }
+#endif
 
 static inline void lock_bin(int i)
 {
diff --git a/libc-top-half/musl/src/prng/random.c b/libc-top-half/musl/src/prng/random.c
index daac028d8..957f2f955 100644
--- a/libc-top-half/musl/src/prng/random.c
+++ b/libc-top-half/musl/src/prng/random.c
@@ -24,8 +24,8 @@ static int i = 3;
 static int j = 0;
 static uint32_t *x = init+1;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-static volatile int lock[1];
-volatile int *const __random_lockptr = lock;
+static __lock_t lock[1];
+__lock_t *const __random_lockptr = lock;
 #endif
 
 static uint32_t lcg31(uint32_t x) {
@@ -67,9 +67,11 @@ static void __srandom(unsigned seed) {
 }
 
 void srandom(unsigned seed) {
-	LOCK(lock);
+	// The random number generator does not hit any context switch points, 
+	// so we can use weak locks here.
+	WEAK_LOCK(lock);
 	__srandom(seed);
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 }
 
 char *initstate(unsigned seed, char *state, size_t size) {
@@ -77,7 +79,7 @@ char *initstate(unsigned seed, char *state, size_t size) {
 
 	if (size < 8)
 		return 0;
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	old = savestate();
 	if (size < 32)
 		n = 0;
@@ -92,24 +94,24 @@ char *initstate(unsigned seed, char *state, size_t size) {
 	x = (uint32_t*)state + 1;
 	__srandom(seed);
 	savestate();
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 	return old;
 }
 
 char *setstate(char *state) {
 	void *old;
 
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	old = savestate();
 	loadstate((uint32_t*)state);
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 	return old;
 }
 
 long random(void) {
 	long k;
 
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	if (n == 0) {
 		k = x[0] = lcg31(x[0]);
 		goto end;
@@ -121,6 +123,6 @@ long random(void) {
 	if (++j == n)
 		j = 0;
 end:
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 	return k;
 }
diff --git a/libc-top-half/musl/src/stdio/__fdopen.c b/libc-top-half/musl/src/stdio/__fdopen.c
index 5c8df495e..205338c2f 100644
--- a/libc-top-half/musl/src/stdio/__fdopen.c
+++ b/libc-top-half/musl/src/stdio/__fdopen.c
@@ -75,7 +75,7 @@ FILE *__fdopen(int fd, const char *mode)
 	f->close = __stdio_close;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	if (!libc.threaded) f->lock = -1;
+	if (!libc.threaded) __STDIO_LOCK_RESET(&f->lock);
 #endif
 
 	/* Add new FILE to open file list */
diff --git a/libc-top-half/musl/src/stdio/__fopen_rb_ca.c b/libc-top-half/musl/src/stdio/__fopen_rb_ca.c
index 192050b00..27c2544fe 100644
--- a/libc-top-half/musl/src/stdio/__fopen_rb_ca.c
+++ b/libc-top-half/musl/src/stdio/__fopen_rb_ca.c
@@ -25,7 +25,7 @@ FILE *__fopen_rb_ca(const char *filename, FILE *f, unsigned char *buf, size_t le
 	f->seek = __stdio_seek;
 	f->close = __stdio_close;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	f->lock = -1;
+	__STDIO_LOCK_RESET(&f->lock);
 #endif
 
 	return f;
diff --git a/libc-top-half/musl/src/stdio/__lockfile.c b/libc-top-half/musl/src/stdio/__lockfile.c
index 0f60a1499..25b19e60c 100644
--- a/libc-top-half/musl/src/stdio/__lockfile.c
+++ b/libc-top-half/musl/src/stdio/__lockfile.c
@@ -1,6 +1,24 @@
 #include "stdio_impl.h"
 #include "pthread_impl.h"
+#include <wasi/version.h>
 
+#ifdef __wasi_cooperative_threads__
+#include "lock.h"
+int __lockfile(FILE *f)
+{
+	// Allow recursive locking
+	if (f->lock.owner == __pthread_self()->tid)
+		return 0;
+
+	LOCK(&f->lock);
+	return 1;
+}
+
+void __unlockfile(FILE *f)
+{
+	UNLOCK(&f->lock);
+}
+#else
 int __lockfile(FILE *f)
 {
 	int owner = f->lock, tid = __pthread_self()->tid;
@@ -21,3 +39,4 @@ void __unlockfile(FILE *f)
 	if (a_swap(&f->lock, 0) & MAYBE_WAITERS)
 		__wake(&f->lock, 1, 1);
 }
+#endif
diff --git a/libc-top-half/musl/src/stdio/fmemopen.c b/libc-top-half/musl/src/stdio/fmemopen.c
index 3ee57b9ea..22628e56f 100644
--- a/libc-top-half/musl/src/stdio/fmemopen.c
+++ b/libc-top-half/musl/src/stdio/fmemopen.c
@@ -130,7 +130,7 @@ FILE *fmemopen(void *restrict buf, size_t size, const char *restrict mode)
 	f->f.close = mclose;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	if (!libc.threaded) f->f.lock = -1;
+	if (!libc.threaded) __STDIO_LOCK_RESET(&f->f.lock);
 #endif
 
 	return __ofl_add(&f->f);
diff --git a/libc-top-half/musl/src/stdio/ftrylockfile.c b/libc-top-half/musl/src/stdio/ftrylockfile.c
index 50650585b..489d46f88 100644
--- a/libc-top-half/musl/src/stdio/ftrylockfile.c
+++ b/libc-top-half/musl/src/stdio/ftrylockfile.c
@@ -4,9 +4,15 @@
 
 void __do_orphaned_stdio_locks()
 {
-	FILE *f;
-	for (f=__pthread_self()->stdio_locks; f; f=f->next_locked)
-		a_store(&f->lock, 0x40000000);
+    FILE *f;
+    for (f=__pthread_self()->stdio_locks; f; f=f->next_locked) {
+		#ifdef __wasi_cooperative_threads__
+        f->lock.owner = 0;
+        __waitlist_wake_all(&f->lock.waiters);
+		#else
+        a_store(&f->lock, 0x40000000);
+		#endif
+    }
 }
 
 void __unlist_locked_file(FILE *f)
@@ -27,6 +33,26 @@ void __register_locked_file(FILE *f, pthread_t self)
 	self->stdio_locks = f;
 }
 
+#ifdef __wasi_cooperative_threads__
+int ftrylockfile(FILE *f)
+{
+	int self_tid = __pthread_self()->tid;
+	if (f->lock.owner == self_tid) {
+		if (f->lockcount == LONG_MAX)
+			return -1;
+		f->lockcount++;
+		return 0;
+	}
+
+	// Try to acquire the lock
+	if (f->lock.owner != 0)
+		return -1;
+
+	f->lock.owner = self_tid;
+	f->lockcount = 1;
+	return 0;
+}
+#else
 int ftrylockfile(FILE *f)
 {
 	pthread_t self = __pthread_self();
@@ -44,3 +70,4 @@ int ftrylockfile(FILE *f)
 	__register_locked_file(f, self);
 	return 0;
 }
+#endif
diff --git a/libc-top-half/musl/src/stdio/getc.h b/libc-top-half/musl/src/stdio/getc.h
index e62e3f0da..61074d607 100644
--- a/libc-top-half/musl/src/stdio/getc.h
+++ b/libc-top-half/musl/src/stdio/getc.h
@@ -1,23 +1,35 @@
 #include "stdio_impl.h"
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 #include "pthread_impl.h"
+#include <wasi/api.h>
 
 #ifdef __GNUC__
 __attribute__((__noinline__))
 #endif
 static int locking_getc(FILE *f)
 {
+#ifdef __wasi_cooperative_threads__
+    __lockfile(f);
+    int c = getc_unlocked(f);
+    __unlockfile(f);
+	return c;
+#else
 	if (a_cas(&f->lock, 0, MAYBE_WAITERS-1)) __lockfile(f);
 	int c = getc_unlocked(f);
 	if (a_swap(&f->lock, 0) & MAYBE_WAITERS)
 		__wake(&f->lock, 1, 1);
 	return c;
+#endif
 }
 #endif
 
 static inline int do_getc(FILE *f)
 {
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+#ifdef __wasi_cooperative_threads__
+	if (f->lock.owner == __pthread_self()->tid)
+		return getc_unlocked(f);
+	return locking_getc(f);
+#elif defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 	int l = f->lock;
 	if (l < 0 || l && (l & ~MAYBE_WAITERS) == __pthread_self()->tid)
 		return getc_unlocked(f);
diff --git a/libc-top-half/musl/src/stdio/ofl.c b/libc-top-half/musl/src/stdio/ofl.c
index 33a8aa50b..d69ee2805 100644
--- a/libc-top-half/musl/src/stdio/ofl.c
+++ b/libc-top-half/musl/src/stdio/ofl.c
@@ -4,8 +4,8 @@
 
 static FILE *ofl_head;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-static volatile int ofl_lock[1];
-volatile int *const __stdio_ofl_lockptr = ofl_lock;
+static __lock_t ofl_lock[1];
+__lock_t *const __stdio_ofl_lockptr = ofl_lock;
 #endif
 
 FILE **__ofl_lock()
diff --git a/libc-top-half/musl/src/stdio/open_memstream.c b/libc-top-half/musl/src/stdio/open_memstream.c
index aa5022065..9690fe8b3 100644
--- a/libc-top-half/musl/src/stdio/open_memstream.c
+++ b/libc-top-half/musl/src/stdio/open_memstream.c
@@ -102,7 +102,7 @@ FILE *open_memstream(char **bufp, size_t *sizep)
 	f->f.mode = -1;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	if (!libc.threaded) f->f.lock = -1;
+	if (!libc.threaded) __STDIO_LOCK_RESET(&f->f.lock);
 #endif
 
 	return __ofl_add(&f->f);
diff --git a/libc-top-half/musl/src/stdio/open_wmemstream.c b/libc-top-half/musl/src/stdio/open_wmemstream.c
index 198d5d439..d06c24db8 100644
--- a/libc-top-half/musl/src/stdio/open_wmemstream.c
+++ b/libc-top-half/musl/src/stdio/open_wmemstream.c
@@ -107,7 +107,7 @@ FILE *open_wmemstream(wchar_t **bufp, size_t *sizep)
 	f->f.close = wms_close;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	if (!libc.threaded) f->f.lock = -1;
+	if (!libc.threaded) __STDIO_LOCK_RESET(&f->f.lock);
 #endif
 
 	fwide(&f->f, 1);
diff --git a/libc-top-half/musl/src/stdio/putc.h b/libc-top-half/musl/src/stdio/putc.h
index 2cc63d2db..7e1d49d6b 100644
--- a/libc-top-half/musl/src/stdio/putc.h
+++ b/libc-top-half/musl/src/stdio/putc.h
@@ -1,23 +1,35 @@
 #include "stdio_impl.h"
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 #include "pthread_impl.h"
+#include <wasi/api.h>
 
 #ifdef __GNUC__
 __attribute__((__noinline__))
 #endif
 static int locking_putc(int c, FILE *f)
 {
+#ifdef __wasi_cooperative_threads__
+	__lockfile(f);
+	c = putc_unlocked(c, f);
+	__unlockfile(f);
+	return c;
+#else
 	if (a_cas(&f->lock, 0, MAYBE_WAITERS-1)) __lockfile(f);
 	c = putc_unlocked(c, f);
 	if (a_swap(&f->lock, 0) & MAYBE_WAITERS)
 		__wake(&f->lock, 1, 1);
 	return c;
+#endif
 }
 #endif
 
 static inline int do_putc(int c, FILE *f)
 {
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+#ifdef __wasi_cooperative_threads__
+	if (f->lock.owner == __pthread_self()->tid)
+		return putc_unlocked(c, f);
+	return locking_putc(c, f);
+#elif defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 	int l = f->lock;
 	if (l < 0 || l && (l & ~MAYBE_WAITERS) == __pthread_self()->tid)
 		return putc_unlocked(c, f);
diff --git a/libc-top-half/musl/src/stdio/stderr.c b/libc-top-half/musl/src/stdio/stderr.c
index 5f24549f8..a14ba5637 100644
--- a/libc-top-half/musl/src/stdio/stderr.c
+++ b/libc-top-half/musl/src/stdio/stderr.c
@@ -13,7 +13,7 @@ hidden FILE __stderr_FILE = {
 	.seek = __stdio_seek,
 	.close = __stdio_close,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	.lock = -1,
+	.lock = __STDIO_LOCK_INIT,
 #endif
 };
 FILE *const stderr = &__stderr_FILE;
diff --git a/libc-top-half/musl/src/stdio/stdin.c b/libc-top-half/musl/src/stdio/stdin.c
index 68e1c3f64..ddc1a892a 100644
--- a/libc-top-half/musl/src/stdio/stdin.c
+++ b/libc-top-half/musl/src/stdio/stdin.c
@@ -12,7 +12,7 @@ hidden FILE __stdin_FILE = {
 	.seek = __stdio_seek,
 	.close = __stdio_close,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	.lock = -1,
+	.lock = __STDIO_LOCK_INIT,
 #endif
 };
 FILE *const stdin = &__stdin_FILE;
diff --git a/libc-top-half/musl/src/stdio/stdout.c b/libc-top-half/musl/src/stdio/stdout.c
index e0e2bced9..914f5a456 100644
--- a/libc-top-half/musl/src/stdio/stdout.c
+++ b/libc-top-half/musl/src/stdio/stdout.c
@@ -13,7 +13,7 @@ hidden FILE __stdout_FILE = {
 	.seek = __stdio_seek,
 	.close = __stdio_close,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	.lock = -1,
+	.lock = __STDIO_LOCK_INIT,
 #endif
 };
 FILE *const stdout = &__stdout_FILE;
diff --git a/libc-top-half/musl/src/stdio/vdprintf.c b/libc-top-half/musl/src/stdio/vdprintf.c
index cef0a1af2..4ab32308a 100644
--- a/libc-top-half/musl/src/stdio/vdprintf.c
+++ b/libc-top-half/musl/src/stdio/vdprintf.c
@@ -6,7 +6,7 @@ int vdprintf(int fd, const char *restrict fmt, va_list ap)
 		.fd = fd, .lbf = EOF, .write = __stdio_write,
 		.buf = (void *)fmt, .buf_size = 0,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-		.lock = -1
+		.lock = __STDIO_LOCK_INIT,
 #endif
 	};
 	return vfprintf(&f, fmt, ap);
diff --git a/libc-top-half/musl/src/stdio/vsnprintf.c b/libc-top-half/musl/src/stdio/vsnprintf.c
index 48c3aef7e..4d8366baf 100644
--- a/libc-top-half/musl/src/stdio/vsnprintf.c
+++ b/libc-top-half/musl/src/stdio/vsnprintf.c
@@ -41,7 +41,7 @@ int vsnprintf(char *restrict s, size_t n, const char *restrict fmt, va_list ap)
 		.lbf = EOF,
 		.write = sn_write,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-		.lock = -1,
+		.lock = __STDIO_LOCK_INIT,
 #endif
 		.buf = buf,
 		.cookie = &c,
diff --git a/libc-top-half/musl/src/stdio/vsscanf.c b/libc-top-half/musl/src/stdio/vsscanf.c
index 0e5b48265..883e1d8c4 100644
--- a/libc-top-half/musl/src/stdio/vsscanf.c
+++ b/libc-top-half/musl/src/stdio/vsscanf.c
@@ -19,10 +19,9 @@ int vsscanf(const char *restrict s, const char *restrict fmt, va_list ap)
 {
 	FILE f = {
 		.buf = (void *)s, .cookie = (void *)s,
+		.read = string_read,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-		.read = string_read, .lock = -1
-#else
-		.read = string_read
+		.lock = __STDIO_LOCK_INIT
 #endif
 	};
 	return vfscanf(&f, fmt, ap);
diff --git a/libc-top-half/musl/src/stdio/vswprintf.c b/libc-top-half/musl/src/stdio/vswprintf.c
index cab94cba7..d17da02ba 100644
--- a/libc-top-half/musl/src/stdio/vswprintf.c
+++ b/libc-top-half/musl/src/stdio/vswprintf.c
@@ -44,7 +44,7 @@ int vswprintf(wchar_t *restrict s, size_t n, const wchar_t *restrict fmt, va_lis
 		.lbf = EOF,
 		.write = sw_write,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-		.lock = -1,
+		.lock = __STDIO_LOCK_INIT,
 #endif
 		.buf = buf,
 		.buf_size = sizeof buf,
diff --git a/libc-top-half/musl/src/stdio/vswscanf.c b/libc-top-half/musl/src/stdio/vswscanf.c
index ea827102d..feef4e451 100644
--- a/libc-top-half/musl/src/stdio/vswscanf.c
+++ b/libc-top-half/musl/src/stdio/vswscanf.c
@@ -30,10 +30,9 @@ int vswscanf(const wchar_t *restrict s, const wchar_t *restrict fmt, va_list ap)
 	FILE f = {
 		.buf = buf, .buf_size = sizeof buf,
 		.cookie = (void *)s,
+		.read = wstring_read,
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-		.read = wstring_read, .lock = -1
-#else
-		.read = wstring_read
+		.lock = __STDIO_LOCK_INIT
 #endif
 	};
 	return vfwscanf(&f, fmt, ap);
diff --git a/libc-top-half/musl/src/stdlib/wcstod.c b/libc-top-half/musl/src/stdlib/wcstod.c
index 97b894eb2..c198f08c5 100644
--- a/libc-top-half/musl/src/stdlib/wcstod.c
+++ b/libc-top-half/musl/src/stdlib/wcstod.c
@@ -44,7 +44,7 @@ static long double wcstox(const wchar_t *s, wchar_t **p, int prec)
 	f.rpos = f.rend = f.buf = buf + 4;
 	f.buf_size = sizeof buf - 4;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	f.lock = -1;
+	__STDIO_LOCK_RESET(&f.lock);
 #endif
 	f.read = do_read;
 	while (iswspace(*t)) t++;
diff --git a/libc-top-half/musl/src/stdlib/wcstol.c b/libc-top-half/musl/src/stdlib/wcstol.c
index 3aefd06f4..43439c0d0 100644
--- a/libc-top-half/musl/src/stdlib/wcstol.c
+++ b/libc-top-half/musl/src/stdlib/wcstol.c
@@ -38,7 +38,7 @@ static unsigned long long wcstox(const wchar_t *s, wchar_t **p, int base, unsign
 	f.rpos = f.rend = f.buf = buf + 4;
 	f.buf_size = sizeof buf - 4;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-	f.lock = -1;
+	__STDIO_LOCK_RESET(&f.lock);
 #endif
 	f.read = do_read;
 	while (iswspace(*t)) t++;
diff --git a/libc-top-half/musl/src/thread/coop-threads/__lock.c b/libc-top-half/musl/src/thread/coop-threads/__lock.c
new file mode 100644
index 000000000..939d87ad9
--- /dev/null
+++ b/libc-top-half/musl/src/thread/coop-threads/__lock.c
@@ -0,0 +1,37 @@
+#include "pthread_impl.h"
+#include "lock.h"
+#include <wasi/api.h>
+
+#ifndef __wasip3__
+#error "Unknown WASI version"
+#endif
+
+void __lock(struct __coop_lock *lock)
+{
+	if (lock->owner == __pthread_self()->tid) {
+		/* Trap on recursive locking. */
+		__builtin_trap();
+	}
+
+	/* Loop until we acquire the lock. */
+	while (lock->owner != 0) {
+		__waitlist_wait_on(&lock->waiters);
+		/* After waking, the lock might still be held by another
+		 * thread that was scheduled before us, so loop back. */
+	}
+	
+	lock->owner = __pthread_self()->tid;
+}
+
+void __unlock(struct __coop_lock *lock)
+{
+	int tid = __pthread_self()->tid;
+	if (lock->owner != tid) {
+		/* We're trying to unlock a lock we don't own. */
+		__builtin_trap();
+	}
+
+	lock->owner = 0;
+	/* Awake one waiter; the others will be resumed on future unlocks. */ 
+	__waitlist_wake_one(&lock->waiters);
+}
\ No newline at end of file
diff --git a/libc-top-half/musl/src/thread/coop-threads/__wait.c b/libc-top-half/musl/src/thread/coop-threads/__wait.c
new file mode 100644
index 000000000..db8646dcc
--- /dev/null
+++ b/libc-top-half/musl/src/thread/coop-threads/__wait.c
@@ -0,0 +1,47 @@
+#include "pthread_impl.h"
+#include <wasi/api.h>
+
+#ifndef __wasip3__
+#error "Unknown WASI version"
+#endif
+
+void __waitlist_wait_on(struct __waitlist_node **list)
+{
+    struct __waitlist_node node = {
+        .tid = __pthread_self()->tid,
+        .next = *list,
+    };
+    *list = &node;
+    
+    wasip3_thread_suspend();
+}
+
+void __waitlist_wake_one(struct __waitlist_node **list)
+{
+    if (*list == NULL) {
+        return;
+    }
+    struct __waitlist_node *node = *list;
+    *list = node->next;
+    wasip3_thread_yield_to_suspended(node->tid);
+}
+
+void __waitlist_wake_all(struct __waitlist_node **list)
+{
+    struct __waitlist_node **prev = list;
+    struct __waitlist_node *curr = *list;
+
+    while (curr) {
+        uint32_t tid = curr->tid;
+        *prev = curr->next;
+        // As a scheduling optimization, we always yield directly to the last
+        // suspended thread instead of just scheduling it to run at some point.
+        if (curr->next == NULL) {
+            wasip3_thread_yield_to_suspended(tid);
+        }
+        else {
+            wasip3_thread_unsuspend(tid);
+        }
+        curr = *prev;
+    }
+}
\ No newline at end of file
diff --git a/libc-top-half/musl/src/thread/wasi-threads/__wait.c b/libc-top-half/musl/src/thread/wasi-threads/__wait.c
index 9484c0f3d..8248571f9 100644
--- a/libc-top-half/musl/src/thread/wasi-threads/__wait.c
+++ b/libc-top-half/musl/src/thread/wasi-threads/__wait.c
@@ -64,3 +64,28 @@ void __wait(volatile int *addr, volatile int *waiters, int val, int priv)
 	}
 	if (waiters) a_dec(waiters);
 }
+
+void __wake(volatile void *addr, int cnt, int priv)
+{
+	if (priv) priv = FUTEX_PRIVATE;
+	if (cnt<0) cnt = INT_MAX;
+#ifdef __wasilibc_unmodified_upstream
+	__syscall(SYS_futex, addr, FUTEX_WAKE|priv, cnt) != -ENOSYS ||
+	__syscall(SYS_futex, addr, FUTEX_WAKE, cnt);
+#else
+#ifdef _REENTRANT
+	__builtin_wasm_memory_atomic_notify((int*)addr, cnt);
+#endif
+#endif
+}
+
+void __futexwait(volatile void *addr, int val, int priv)
+{
+#ifdef __wasilibc_unmodified_upstream
+	if (priv) priv = FUTEX_PRIVATE;
+	__syscall(SYS_futex, addr, FUTEX_WAIT|priv, val, 0) != -ENOSYS ||
+	__syscall(SYS_futex, addr, FUTEX_WAIT, val, 0);
+#else
+	__wait(addr, NULL, val, priv);
+#endif
+}
\ No newline at end of file
diff --git a/libc-top-half/musl/src/thread/wasi-threads/pthread_create.c b/libc-top-half/musl/src/thread/wasi-threads/pthread_create.c
index 879b48c1f..4d1c02f16 100644
--- a/libc-top-half/musl/src/thread/wasi-threads/pthread_create.c
+++ b/libc-top-half/musl/src/thread/wasi-threads/pthread_create.c
@@ -349,7 +349,9 @@ weak_alias(dummy_file, __stderr_used);
 
 static void init_file_lock(FILE *f)
 {
+#ifndef __wasi_cooperative_threads__
 	if (f && f->lock<0) f->lock = 0;
+#endif
 }
 
 int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index bb51b2c05..d24aeec96 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -126,6 +126,7 @@ function(register_test test_name executable_name)
   endif()
   if (WASI STREQUAL "p3")
     list(APPEND wasmtime_args --wasm component-model-async)
+    list(APPEND wasmtime_args --wasm component-model-threading)
     list(APPEND wasmtime_args --wasi p3)
   endif()