Deadlock in pthread_barrier_wait #402
Description
Hi, I work on support of WASI-threads in WAMR.
I discovered a reproducible situation where 4 threads infinitely hang on pthread_barrier_wait.
Below I attach the minimal sample that hangs, it's reproducible if test run 500+ times. Left only calls to pthread_barrier_wait so we are sure that it doesn't stuck anywhere else.
#include <assert.h>
#include <pthread.h>
#include <malloc.h>
#define NUM_THREADS 3
static pthread_barrier_t barrier;
#define STACK_SIZE 32 * 1024 // same as the main stack
typedef struct {
void *stack;
} start_args_t;
static inline int
start_args_init(start_args_t *start_args)
{
start_args->stack = malloc(STACK_SIZE);
if (!start_args->stack) {
return 0;
}
start_args->stack += STACK_SIZE;
return 1;
}
static inline void
start_args_deinit(start_args_t *start_args)
{
free(start_args->stack - STACK_SIZE);
}
void
__wasi_thread_start_C(int thread_id, int *start_arg)
{
pthread_barrier_wait(&barrier); <------ 3 threads stuck here and wasm_runtime_atomic_wait called inside
}
int
main(int argc, char **argv)
{
assert(pthread_barrier_init(&barrier, NULL, NUM_THREADS + 1) == 0
&& "Failed to init barrier");
int i;
int thread_id = -1;
start_args_t data[NUM_THREADS] = { 0 };
for (i = 0; i < NUM_THREADS; i++) {
/* No graceful memory free to simplify the test */
assert(start_args_init(&data[i])
&& "Failed to allocate thread's stack");
}
for (i = 0; i < NUM_THREADS; i++) {
thread_id = __wasi_thread_spawn(&data[i]);
assert(thread_id > 0 && "Failed to create thread");
}
pthread_barrier_wait(&barrier); <------ main thread stuck here and wasm_runtime_atomic_wait called inside
}
I used wasi-sdk-20+threads to compile .wasm file from this test with this command
/path/to/wasi-sdk/bin/clang \
-g \
--sysroot /path/to/wasi-sdk/share/wasi-sysroot/ \
-target wasm32-wasi-threads \
-pthread -ftls-model=local-exec \
-z stack-size=32768 \
-Wl,--export=__heap_base \
-Wl,--export=__data_end \
-Wl,--import-memory,--export-memory,--max-memory=1048576 \
-Wl,--shared-memory,--max-memory=1966080 \
-Wl,--export=wasi_thread_start \
-Wl,--export=malloc \
-Wl,--export=free \
test.c -o test.wasm
To build WAMR that support threads this short manual can be followed.
Before compiling I recommend to change this value to 1e3 or lower so it's way faster to reproduce.
After WAMR(iwasm) and test.wasm are built I ran
for i in `seq 5000`; do ./iwasm -v=5 /tmp/hang_min.wasm ; done