struct pthread_queue ready_queue = TAILQ_HEAD_INITIALIZER(ready_queue);
struct pthread_queue active_queue = TAILQ_HEAD_INITIALIZER(active_queue);
-mcs_lock_t queue_lock = MCS_LOCK_INIT;
+struct mcs_pdr_lock queue_lock;
pthread_once_t init_once = PTHREAD_ONCE_INIT;
int threads_ready = 0;
int threads_active = 0;
static inline void spin_to_sleep(unsigned int spins, unsigned int *spun);
/* Pthread 2LS operations */
-struct uthread *pth_init(void);
void pth_sched_entry(void);
-struct uthread *pth_thread_create(void (*func)(void), void *udata);
void pth_thread_runnable(struct uthread *uthread);
void pth_thread_yield(struct uthread *uthread);
-void pth_thread_destroy(struct uthread *uthread);
+void pth_thread_paused(struct uthread *uthread);
void pth_preempt_pending(void);
void pth_spawn_thread(uintptr_t pc_start, void *data);
void pth_blockon_sysc(struct syscall *sysc);
static void pth_handle_syscall(struct event_msg *ev_msg, unsigned int ev_type);
struct schedule_ops pthread_sched_ops = {
- pth_init,
pth_sched_entry,
- pth_thread_create,
pth_thread_runnable,
pth_thread_yield,
- pth_thread_destroy,
+ pth_thread_paused,
pth_blockon_sysc,
0, /* pth_preempt_pending, */
0, /* pth_spawn_thread, */
static void __pthread_free_stack(struct pthread_tcb *pt);
static int __pthread_allocate_stack(struct pthread_tcb *pt);
-/* Do whatever init you want. Return a uthread representing thread0 (int
- * main()) */
-struct uthread *pth_init(void)
-{
- uintptr_t mmap_block;
- struct mcs_lock_qnode local_qn = {0};
- /* Tell the kernel where and how we want to receive events. This is just an
- * example of what to do to have a notification turned on. We're turning on
- * USER_IPIs, posting events to vcore 0's vcpd, and telling the kernel to
- * send to vcore 0. Note sys_self_notify will ignore the vcoreid pref.
- * Also note that enable_kevent() is just an example, and you probably want
- * to use parts of event.c to do what you want. */
- enable_kevent(EV_USER_IPI, 0, EVENT_IPI);
-
- /* Handle syscall events. Using small ev_qs, with no internal ev_mbox. */
- ev_handlers[EV_SYSCALL] = pth_handle_syscall;
- /* Set up the per-vcore structs to track outstanding syscalls */
- sysc_mgmt = malloc(sizeof(struct sysc_mgmt) * max_vcores());
- assert(sysc_mgmt);
-#if 1 /* Independent ev_mboxes per vcore */
- /* Get a block of pages for our per-vcore (but non-VCPD) ev_qs */
- mmap_block = (uintptr_t)mmap(0, PGSIZE * 2 * max_vcores(),
- PROT_WRITE | PROT_READ,
- MAP_POPULATE | MAP_ANONYMOUS, -1, 0);
- assert(mmap_block);
- /* Could be smarter and do this on demand (in case we don't actually want
- * max_vcores()). */
- for (int i = 0; i < max_vcores(); i++) {
- /* Each vcore needs to point to a non-VCPD ev_q */
- sysc_mgmt[i].ev_q = get_big_event_q_raw();
- sysc_mgmt[i].ev_q->ev_flags = EVENT_IPI; /* totally up to you */
- sysc_mgmt[i].ev_q->ev_vcore = i;
- ucq_init_raw(&sysc_mgmt[i].ev_q->ev_mbox->ev_msgs,
- mmap_block + (2 * i ) * PGSIZE,
- mmap_block + (2 * i + 1) * PGSIZE);
- }
- /* Technically, we should munmap and free what we've alloc'd, but the
- * kernel will clean it up for us when we exit. */
-#endif
-#if 0 /* One global ev_mbox, separate ev_q per vcore */
- struct event_mbox *sysc_mbox = malloc(sizeof(struct event_mbox));
- uintptr_t two_pages = (uintptr_t)mmap(0, PGSIZE * 2, PROT_WRITE | PROT_READ,
- MAP_POPULATE | MAP_ANONYMOUS, -1, 0);
- printd("Global ucq: %08p\n", &sysc_mbox->ev_msgs);
- assert(sysc_mbox);
- assert(two_pages);
- memset(sysc_mbox, 0, sizeof(struct event_mbox));
- ucq_init_raw(&sysc_mbox->ev_msgs, two_pages, two_pages + PGSIZE);
- for (int i = 0; i < max_vcores(); i++) {
- sysc_mgmt[i].ev_q = get_event_q();
- sysc_mgmt[i].ev_q->ev_flags = EVENT_IPI;
- sysc_mgmt[i].ev_q->ev_vcore = i;
- sysc_mgmt[i].ev_q->ev_mbox = sysc_mbox;
- }
-#endif
-
- /* Create a pthread_tcb for the main thread */
- pthread_t t = (pthread_t)calloc(1, sizeof(struct pthread_tcb));
- assert(t);
- t->id = get_next_pid();
- t->stacksize = USTACK_NUM_PAGES * PGSIZE;
- t->stacktop = (void*)USTACKTOP;
- t->detached = TRUE;
- t->flags = 0;
- t->finished = 0;
- assert(t->id == 0);
- /* Put the new pthread on the active queue */
- mcs_lock_notifsafe(&queue_lock, &local_qn);
- threads_active++;
- TAILQ_INSERT_TAIL(&active_queue, t, next);
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
- return (struct uthread*)t;
-}
-
/* Called from vcore entry. Options usually include restarting whoever was
* running there before or running a new thread. Events are handled out of
* event.c (table of function pointers, stuff like that). */
}
/* no one currently running, so lets get someone from the ready queue */
struct pthread_tcb *new_thread = NULL;
- struct mcs_lock_qnode local_qn = {0};
- /* For now, let's spin and handle events til we get a thread to run. This
- * will help catch races, instead of only having one core ever run a thread
- * (if there is just one, etc). Also, we don't need the EVENT_IPIs for this
- * to work (since we poll handle_events() */
- while (!new_thread) {
+ /* Try to get a thread. If we get one, we'll break out and run it. If not,
+ * we'll try to yield. vcore_yield() might return, if we lost a race and
+ * had a new event come in, one that may make us able to get a new_thread */
+ do {
handle_events(vcoreid);
- mcs_lock_notifsafe(&queue_lock, &local_qn);
+ __check_preempt_pending(vcoreid);
+ mcs_pdr_lock(&queue_lock);
new_thread = TAILQ_FIRST(&ready_queue);
if (new_thread) {
TAILQ_REMOVE(&ready_queue, new_thread, next);
TAILQ_INSERT_TAIL(&active_queue, new_thread, next);
threads_active++;
threads_ready--;
+ mcs_pdr_unlock(&queue_lock);
+ /* If you see what looks like the same uthread running in multiple
+ * places, your list might be jacked up. Turn this on. */
+ printd("[P] got uthread %08p on vc %d state %08p flags %08p\n",
+ new_thread, vcoreid,
+ ((struct uthread*)new_thread)->state,
+ ((struct uthread*)new_thread)->flags);
+ break;
}
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
- }
- /* Instead of yielding, you could spin, turn off the core, set an alarm,
- * whatever. You want some logic to decide this. Uthread code wil have
- * helpers for this (like how we provide run_uthread()) */
- if (!new_thread) {
- /* Note, we currently don't get here (due to the while loop) */
+ mcs_pdr_unlock(&queue_lock);
+ /* no new thread, try to yield */
printd("[P] No threads, vcore %d is yielding\n", vcore_id());
- /* Not actually yielding - just spin for now, so we can get syscall
- * unblocking events */
- vcore_idle();
- //sys_yield(0);
- assert(0);
- }
+ /* TODO: you can imagine having something smarter here, like spin for a
+ * bit before yielding (or not at all if you want to be greedy). */
+ vcore_yield(FALSE);
+ } while (1);
assert(((struct uthread*)new_thread)->state != UT_RUNNING);
run_uthread((struct uthread*)new_thread);
assert(0);
pthread_exit(me->start_routine(me->arg));
}
-/* Responible for creating the uthread and initializing its user trap frame */
-struct uthread *pth_thread_create(void (*func)(void), void *udata)
-{
- struct pthread_tcb *pthread;
- pthread_attr_t *attr = (pthread_attr_t*)udata;
- pthread = (pthread_t)calloc(1, sizeof(struct pthread_tcb));
- assert(pthread);
- pthread->stacksize = PTHREAD_STACK_SIZE; /* default */
- pthread->finished = 0;
- pthread->flags = 0;
- pthread->id = get_next_pid();
- pthread->detached = FALSE; /* default */
- /* Respect the attributes */
- if (attr) {
- if (attr->stacksize) /* don't set a 0 stacksize */
- pthread->stacksize = attr->stacksize;
- if (attr->detachstate == PTHREAD_CREATE_DETACHED)
- pthread->detached = TRUE;
- }
- /* allocate a stack */
- if (__pthread_allocate_stack(pthread))
- printf("We're fucked\n");
- /* Set the u_tf to start up in __pthread_run, which will call the real
- * start_routine and pass it the arg. Note those aren't set until later in
- * pthread_create(). */
- init_user_tf(&pthread->uthread.utf, (uint32_t)__pthread_run,
- (uint32_t)(pthread->stacktop));
- return (struct uthread*)pthread;
-}
-
void pth_thread_runnable(struct uthread *uthread)
{
struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
- struct mcs_lock_qnode local_qn = {0};
/* Insert the newly created thread into the ready queue of threads.
* It will be removed from this queue later when vcore_entry() comes up */
- mcs_lock_notifsafe(&queue_lock, &local_qn);
+ mcs_pdr_lock(&queue_lock);
TAILQ_INSERT_TAIL(&ready_queue, pthread, next);
threads_ready++;
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
+ mcs_pdr_unlock(&queue_lock);
/* Smarter schedulers should look at the num_vcores() and how much work is
* going on to make a decision about how many vcores to request. */
vcore_request(threads_ready);
void pth_thread_yield(struct uthread *uthread)
{
struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
- struct mcs_lock_qnode local_qn = {0};
- /* Remove from the active list, whether exiting or yielding. We're holding
- * the lock throughout both list modifications (if applicable). */
- mcs_lock_notifsafe(&queue_lock, &local_qn);
+ struct pthread_tcb *temp_pth = 0; /* used for exiting AND joining */
+ /* Remove from the active list, whether exiting or yielding. */
+ mcs_pdr_lock(&queue_lock);
threads_active--;
TAILQ_REMOVE(&active_queue, pthread, next);
+ mcs_pdr_unlock(&queue_lock);
if (pthread->flags & PTHREAD_EXITING) {
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
- uthread_destroy(uthread);
+ /* Destroy the pthread */
+ uthread_cleanup(uthread);
+ /* Cleanup, mirroring pthread_create() */
+ __pthread_free_stack(pthread);
+ /* TODO: race on detach state */
+ if (pthread->detached) {
+ free(pthread);
+ } else {
+ /* See if someone is joining on us. If not, we're done (and the
+ * joiner will wake itself when it saw us there instead of 0). */
+ temp_pth = atomic_swap_ptr((void**)&pthread->joiner, pthread);
+ if (temp_pth) {
+ /* they joined before we exited, we need to wake them */
+ printd("[pth] %08p exiting, waking joiner %08p\n",
+ pthread, temp_pth);
+ pth_thread_runnable((struct uthread*)temp_pth);
+ }
+ }
+ } else if (pthread->flags & PTHREAD_JOINING) {
+ /* We're trying to join, yield til we get woken up */
+ /* put ourselves in the join target's joiner slot. If we get anything
+ * back, we lost the race and need to wake ourselves. */
+ temp_pth = atomic_swap_ptr((void**)&pthread->join_target->joiner,
+ pthread);
+ /* after that atomic swap, the pthread might be woken up (if it
+ * succeeded), so don't touch pthread again after that (this following
+ * if () is okay). */
+ if (temp_pth) {
+ assert(temp_pth == pthread->join_target); /* Sanity */
+ /* wake ourselves, not the exited one! */
+ printd("[pth] %08p already exit, rewaking ourselves, joiner %08p\n",
+ temp_pth, pthread);
+ pth_thread_runnable((struct uthread*)pthread);
+ }
} else {
- /* Put it on the ready list (tail). Don't do this until we are done
- * completely with the thread, since it can be restarted somewhere else.
- * */
- threads_ready++;
- TAILQ_INSERT_TAIL(&ready_queue, pthread, next);
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
+ /* Yielding for no apparent reason (being nice / help break deadlocks).
+ * Just wake it up and make it ready again. */
+ pth_thread_runnable((struct uthread*)pthread);
}
}
-
-void pth_thread_destroy(struct uthread *uthread)
+
+/* For some reason not under its control, the uthread stopped running (compared
+ * to yield, which was caused by uthread/2LS code).
+ *
+ * The main case for this is if the vcore was preempted or if the vcore it was
+ * running on needed to stop. You are given a uthread that looks like it took a
+ * notif, and had its context/silly state copied out to the uthread struct.
+ * (copyout_uthread). Note that this will be called in the context (TLS) of the
+ * vcore that is losing the uthread. If that vcore is running, it'll be in a
+ * preempt-event handling loop (not in your 2LS code). If this is a big
+ * problem, I'll change it. */
+void pth_thread_paused(struct uthread *uthread)
{
struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
- /* Cleanup, mirroring pth_thread_create() */
- __pthread_free_stack(pthread);
- /* TODO: race on detach state */
- if (pthread->detached)
- free(pthread);
- else
- pthread->finished = 1;
+ /* Remove from the active list. Note that I don't particularly care about
+ * the active list. We keep it around because it causes bugs and keeps us
+ * honest. After all, some 2LS may want an active list */
+ mcs_pdr_lock(&queue_lock);
+ threads_active--;
+ TAILQ_REMOVE(&active_queue, pthread, next);
+ mcs_pdr_unlock(&queue_lock);
+ /* At this point, you could do something clever, like put it at the front of
+ * the runqueue, see if it was holding a lock, do some accounting, or
+ * whatever. */
+ uthread_runnable(uthread);
}
void pth_preempt_pending(void)
assert(current_uthread->state == UT_BLOCKED);
/* rip from the active queue */
- struct mcs_lock_qnode local_qn = {0};
struct pthread_tcb *pthread = (struct pthread_tcb*)current_uthread;
- mcs_lock_notifsafe(&queue_lock, &local_qn);
+ mcs_pdr_lock(&queue_lock);
threads_active--;
TAILQ_REMOVE(&active_queue, pthread, next);
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
+ mcs_pdr_unlock(&queue_lock);
/* Set things up so we can wake this thread up later */
sysc->u_data = current_uthread;
return 0;
}
-int pthread_create(pthread_t* thread, const pthread_attr_t* attr,
- void *(*start_routine)(void *), void* arg)
+/* Do whatever init you want. At some point call uthread_lib_init() and pass it
+ * a uthread representing thread0 (int main()) */
+static int pthread_lib_init(void)
{
- struct pthread_tcb *pthread =
- (struct pthread_tcb*)uthread_create(__pthread_run, (void*)attr);
- if (!pthread)
+ /* Make sure this only runs once */
+ static bool initialized = FALSE;
+ if (initialized)
return -1;
+ initialized = TRUE;
+ uintptr_t mmap_block;
+ mcs_pdr_init(&queue_lock);
+ /* Create a pthread_tcb for the main thread */
+ pthread_t t = (pthread_t)calloc(1, sizeof(struct pthread_tcb));
+ assert(t);
+ t->id = get_next_pid();
+ t->stacksize = USTACK_NUM_PAGES * PGSIZE;
+ t->stacktop = (void*)USTACKTOP;
+ t->detached = TRUE;
+ t->flags = 0;
+ t->join_target = 0;
+ t->joiner = 0;
+ assert(t->id == 0);
+ /* Put the new pthread (thread0) on the active queue */
+ mcs_pdr_lock(&queue_lock); /* arguably, we don't need these (_S mode) */
+ threads_active++;
+ TAILQ_INSERT_TAIL(&active_queue, t, next);
+ mcs_pdr_unlock(&queue_lock);
+ /* Tell the kernel where and how we want to receive events. This is just an
+ * example of what to do to have a notification turned on. We're turning on
+ * USER_IPIs, posting events to vcore 0's vcpd, and telling the kernel to
+ * send to vcore 0. Note sys_self_notify will ignore the vcoreid and
+ * private preference. Also note that enable_kevent() is just an example,
+ * and you probably want to use parts of event.c to do what you want. */
+ enable_kevent(EV_USER_IPI, 0, EVENT_IPI | EVENT_VCORE_PRIVATE);
+
+ /* Handle syscall events. */
+ ev_handlers[EV_SYSCALL] = pth_handle_syscall;
+ /* Set up the per-vcore structs to track outstanding syscalls */
+ sysc_mgmt = malloc(sizeof(struct sysc_mgmt) * max_vcores());
+ assert(sysc_mgmt);
+#if 1 /* Independent ev_mboxes per vcore */
+ /* Get a block of pages for our per-vcore (but non-VCPD) ev_qs */
+ mmap_block = (uintptr_t)mmap(0, PGSIZE * 2 * max_vcores(),
+ PROT_WRITE | PROT_READ,
+ MAP_POPULATE | MAP_ANONYMOUS, -1, 0);
+ assert(mmap_block);
+ /* Could be smarter and do this on demand (in case we don't actually want
+ * max_vcores()). */
+ for (int i = 0; i < max_vcores(); i++) {
+ /* Each vcore needs to point to a non-VCPD ev_q */
+ sysc_mgmt[i].ev_q = get_big_event_q_raw();
+ sysc_mgmt[i].ev_q->ev_flags = EVENT_IPI | EVENT_INDIR | EVENT_FALLBACK;
+ sysc_mgmt[i].ev_q->ev_vcore = i;
+ ucq_init_raw(&sysc_mgmt[i].ev_q->ev_mbox->ev_msgs,
+ mmap_block + (2 * i ) * PGSIZE,
+ mmap_block + (2 * i + 1) * PGSIZE);
+ }
+ /* Technically, we should munmap and free what we've alloc'd, but the
+ * kernel will clean it up for us when we exit. */
+#endif
+#if 0 /* One global ev_mbox, separate ev_q per vcore */
+ struct event_mbox *sysc_mbox = malloc(sizeof(struct event_mbox));
+ uintptr_t two_pages = (uintptr_t)mmap(0, PGSIZE * 2, PROT_WRITE | PROT_READ,
+ MAP_POPULATE | MAP_ANONYMOUS, -1, 0);
+ printd("Global ucq: %08p\n", &sysc_mbox->ev_msgs);
+ assert(sysc_mbox);
+ assert(two_pages);
+ memset(sysc_mbox, 0, sizeof(struct event_mbox));
+ ucq_init_raw(&sysc_mbox->ev_msgs, two_pages, two_pages + PGSIZE);
+ for (int i = 0; i < max_vcores(); i++) {
+ sysc_mgmt[i].ev_q = get_event_q();
+ sysc_mgmt[i].ev_q->ev_flags = EVENT_IPI | EVENT_INDIR | EVENT_FALLBACK;
+ sysc_mgmt[i].ev_q->ev_vcore = i;
+ sysc_mgmt[i].ev_q->ev_mbox = sysc_mbox;
+ }
+#endif
+ /* Initialize the uthread code (we're in _M mode after this). Doing this
+ * last so that all the event stuff is ready when we're in _M mode. Not a
+ * big deal one way or the other. Note that vcore_init() hasn't happened
+ * yet, so if a 2LS somehow wants to have its init stuff use things like
+ * vcore stacks or TLSs, we'll need to change this. */
+ assert(!uthread_lib_init((struct uthread*)t));
+ return 0;
+}
+
+int pthread_create(pthread_t *thread, const pthread_attr_t *attr,
+ void *(*start_routine)(void *), void *arg)
+{
+ static bool first = TRUE;
+ if (first) {
+ assert(!pthread_lib_init());
+ first = FALSE;
+ }
+ /* Create the actual thread */
+ struct pthread_tcb *pthread;
+ pthread = (pthread_t)calloc(1, sizeof(struct pthread_tcb));
+ assert(pthread);
+ pthread->stacksize = PTHREAD_STACK_SIZE; /* default */
+ pthread->flags = 0;
+ pthread->id = get_next_pid();
+ pthread->detached = FALSE; /* default */
+ pthread->join_target = 0;
+ pthread->joiner = 0;
+ /* Respect the attributes */
+ if (attr) {
+ if (attr->stacksize) /* don't set a 0 stacksize */
+ pthread->stacksize = attr->stacksize;
+ if (attr->detachstate == PTHREAD_CREATE_DETACHED)
+ pthread->detached = TRUE;
+ }
+ /* allocate a stack */
+ if (__pthread_allocate_stack(pthread))
+ printf("We're fucked\n");
+ /* Set the u_tf to start up in __pthread_run, which will call the real
+ * start_routine and pass it the arg. Note those aren't set until later in
+ * pthread_create(). */
+ init_user_tf(&pthread->uthread.utf, (long)&__pthread_run,
+ (long)(pthread->stacktop));
pthread->start_routine = start_routine;
pthread->arg = arg;
+ /* Initialize the uthread */
+ uthread_init((struct uthread*)pthread);
uthread_runnable((struct uthread*)pthread);
*thread = pthread;
return 0;
int pthread_join(pthread_t thread, void** retval)
{
+ struct pthread_tcb *caller = (struct pthread_tcb*)current_uthread;
/* Not sure if this is the right semantics. There is a race if we deref
* thread and he is already freed (which would have happened if he was
* detached. */
printf("[pthread] trying to join on a detached pthread");
return -1;
}
- while (!thread->finished)
- pthread_yield();
+ /* See if it is already done, to avoid the pain of a uthread_yield() (the
+ * early check is an optimization, pth_thread_yield() handles the race). */
+ if (!thread->joiner) {
+ /* Time to join, set things up so pth_thread_yield() knows what to do */
+ caller->flags |= PTHREAD_JOINING;
+ caller->join_target = thread;
+ uthread_yield(TRUE);
+ /* When we return/restart, the thread will be done */
+ } else {
+ assert(thread->joiner == thread); /* sanity check */
+ }
if (retval)
*retval = thread->retval;
free(thread);
cpu_relax();
spin_to_sleep(PTHREAD_MUTEX_SPINS, &spinner);
}
+ /* normally we'd need a wmb() and a wrmb() after locking, but the
+ * atomic_swap handles the CPU mb(), so just a cmb() is necessary. */
+ cmb();
return 0;
}
int pthread_mutex_unlock(pthread_mutex_t* m)
{
- /* Need to prevent the compiler (and some arches) from reordering older
- * stores */
+ /* keep reads and writes inside the protected region */
+ rwmb();
wmb();
atomic_set(&m->lock, 0);
return 0;
projects / akaros.git / blobdiff