unix: Implement garbage collection with threading.

This patch allows any given thread to do a proper garbage collection and
scan all the pointers of all active threads.

unix/gccollect.c

@@ -136,15 +136,22 @@ STATIC void gc_helper_get_regs(regs_t arr) {
 
 #endif // MICROPY_GCREGS_SETJMP
 
-void gc_collect(void) {
+void gc_collect_regs_and_stack(void) {
-    //gc_dump_info();
-
-    gc_collect_start();
     regs_t regs;
     gc_helper_get_regs(regs);
     // GC stack (and regs because we captured them)
     void **regs_ptr = (void**)(void*)&regs;
     gc_collect_root(regs_ptr, ((uintptr_t)MP_STATE_THREAD(stack_top) - (uintptr_t)&regs) / sizeof(uintptr_t));
+}
+
+void gc_collect(void) {
+    //gc_dump_info();
+
+    gc_collect_start();
+    gc_collect_regs_and_stack();
+    #if MICROPY_PY_THREAD
+    mp_thread_gc_others();
+    #endif
     #if MICROPY_EMIT_NATIVE
     mp_unix_mark_exec();
     #endif

unix/mpthreadport.c

@@ -24,18 +24,88 @@
  * THE SOFTWARE.
  */
 
+#include <stdio.h>
+#include <stdlib.h>
 #include <errno.h>
 
 #include "py/mpstate.h"
 #include "py/mpthread.h"
+#include "py/gc.h"
 
 #if MICROPY_PY_THREAD
 
+#include <signal.h>
+#include <sched.h>
+
+// this structure forms a linked list, one node per active thread
+typedef struct _thread_t {
+    pthread_t id;           // system id of thread
+    int ready;              // whether the thread is ready and running
+    void *arg;              // thread Python args, a GC root pointer
+    struct _thread_t *next;
+} thread_t;
+
 STATIC pthread_key_t tls_key;
 
+// the mutex controls access to the linked list
+STATIC pthread_mutex_t thread_mutex = PTHREAD_MUTEX_INITIALIZER;
+STATIC thread_t *thread;
+
+// this is used to synchronise the signal handler of the thread
+// it's needed because we can't use any pthread calls in a signal handler
+STATIC volatile int thread_signal_done;
+
+// this signal handler is used to scan the regs and stack of a thread
+STATIC void mp_thread_gc(int signo) {
+    if (signo == SIGUSR1) {
+        void gc_collect_regs_and_stack(void);
+        gc_collect_regs_and_stack();
+        thread_signal_done = 1;
+    }
+}
+
 void mp_thread_init(void) {
     pthread_key_create(&tls_key, NULL);
     pthread_setspecific(tls_key, &mp_state_ctx.thread);
+
+    // create first entry in linked list of all threads
+    thread = malloc(sizeof(thread_t));
+    thread->id = pthread_self();
+    thread->ready = 1;
+    thread->arg = NULL;
+    thread->next = NULL;
+
+    // enable signal handler for garbage collection
+    struct sigaction sa;
+    sa.sa_flags = 0;
+    sa.sa_handler = mp_thread_gc;
+    sigemptyset(&sa.sa_mask);
+    sigaction(SIGUSR1, &sa, NULL);
+}
+
+// This function scans all pointers that are external to the current thread.
+// It does this by signalling all other threads and getting them to scan their
+// own registers and stack.  Note that there may still be some edge cases left
+// with race conditions and root-pointer scanning: a given thread may manipulate
+// the global root pointers (in mp_state_ctx) while another thread is doing a
+// garbage collection and tracing these pointers.
+void mp_thread_gc_others(void) {
+    pthread_mutex_lock(&thread_mutex);
+    for (thread_t *th = thread; th != NULL; th = th->next) {
+        gc_collect_root(&th->arg, 1);
+        if (th->id == pthread_self()) {
+            continue;
+        }
+        if (!th->ready) {
+            continue;
+        }
+        thread_signal_done = 0;
+        pthread_kill(th->id, SIGUSR1);
+        while (thread_signal_done == 0) {
+            sched_yield();
+        }
+    }
+    pthread_mutex_unlock(&thread_mutex);
 }
 
 mp_state_thread_t *mp_thread_get_state(void) {
@@ -46,6 +116,17 @@ void mp_thread_set_state(void *state) {
     pthread_setspecific(tls_key, state);
 }
 
+void mp_thread_start(void) {
+    pthread_mutex_lock(&thread_mutex);
+    for (thread_t *th = thread; th != NULL; th = th->next) {
+        if (th->id == pthread_self()) {
+            th->ready = 1;
+            break;
+        }
+    }
+    pthread_mutex_unlock(&thread_mutex);
+}
+
 void mp_thread_create(void *(*entry)(void*), void *arg, size_t stack_size) {
     // default stack size is 8k machine-words
     if (stack_size == 0) {
@@ -63,19 +144,44 @@ void mp_thread_create(void *(*entry)(void*), void *arg, size_t stack_size) {
         goto er;
     }
 
+    pthread_mutex_lock(&thread_mutex);
+
     // create thread
     pthread_t id;
     ret = pthread_create(&id, &attr, entry, arg);
     if (ret != 0) {
+        pthread_mutex_unlock(&thread_mutex);
         goto er;
     }
 
+    // add thread to linked list of all threads
+    thread_t *th = malloc(sizeof(thread_t));
+    th->id = id;
+    th->ready = 0;
+    th->arg = arg;
+    th->next = thread;
+    thread = th;
+
+    pthread_mutex_unlock(&thread_mutex);
+
     return;
 
 er:
     nlr_raise(mp_obj_new_exception_arg1(&mp_type_OSError, MP_OBJ_NEW_SMALL_INT(ret)));
 }
 
+void mp_thread_finish(void) {
+    pthread_mutex_lock(&thread_mutex);
+    // TODO unlink from list
+    for (thread_t *th = thread; th != NULL; th = th->next) {
+        if (th->id == pthread_self()) {
+            th->ready = 0;
+            break;
+        }
+    }
+    pthread_mutex_unlock(&thread_mutex);
+}
+
 void mp_thread_mutex_init(mp_thread_mutex_t *mutex) {
     pthread_mutex_init(mutex, NULL);
 }

unix/mpthreadport.h

@@ -31,5 +31,6 @@
 typedef pthread_mutex_t mp_thread_mutex_t;
 
 void mp_thread_init(void);
+void mp_thread_gc_others(void);
 
 #endif // __MICROPY_INCLUDED_UNIX_MPTHREADPORT_H__