-#include <ros/arch/trapframe.h>
+#include <ros/trapframe.h>
#include <pthread.h>
#include <vcore.h>
#include <mcs.h>
#include <arch/arch.h>
#include <sys/queue.h>
#include <sys/mman.h>
-#include <assert.h>
#include <event.h>
#include <ucq.h>
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;
-pthread_once_t init_once = PTHREAD_ONCE_INIT;
+struct mcs_pdr_lock queue_lock;
int threads_ready = 0;
int threads_active = 0;
+bool can_adjust_vcores = TRUE;
/* Array of per-vcore structs to manage waiting on syscalls and handling
* overflow. Init'd in pth_init(). */
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_thread_blockon_sysc(struct uthread *uthread, void *sysc);
+void pth_thread_has_blocked(struct uthread *uthread, int flags);
void pth_preempt_pending(void);
void pth_spawn_thread(uintptr_t pc_start, void *data);
-void pth_blockon_sysc(struct syscall *sysc);
/* Event Handlers */
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_blockon_sysc,
+ pth_thread_paused,
+ pth_thread_blockon_sysc,
+ pth_thread_has_blocked,
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 | 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
-
- /* 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};
/* 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_unlock_notifsafe(&queue_lock, &local_qn);
+ 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);
+ mcs_pdr_unlock(&queue_lock);
/* no new thread, try to yield */
printd("[P] No threads, vcore %d is yielding\n", vcore_id());
/* 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();
+ if (can_adjust_vcores)
+ vcore_yield(FALSE);
} while (1);
- assert(((struct uthread*)new_thread)->state != UT_RUNNING);
+ assert(new_thread->state == PTH_RUNNABLE);
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;
-}
-
+/* GIANT WARNING: if you make any changes to this, also change the broadcast
+ * wakeups (cond var, barrier, etc) */
void pth_thread_runnable(struct uthread *uthread)
{
struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
- struct mcs_lock_qnode local_qn = {0};
+ /* At this point, the 2LS can see why the thread blocked and was woken up in
+ * the first place (coupling these things together). On the yield path, the
+ * 2LS was involved and was able to set the state. Now when we get the
+ * thread back, we can take a look. */
+ printd("pthread %08p runnable, state was %d\n", pthread, pthread->state);
+ switch (pthread->state) {
+ case (PTH_CREATED):
+ case (PTH_BLK_YIELDING):
+ case (PTH_BLK_JOINING):
+ case (PTH_BLK_SYSC):
+ case (PTH_BLK_PAUSED):
+ case (PTH_BLK_MUTEX):
+ /* can do whatever for each of these cases */
+ break;
+ default:
+ printf("Odd state %d for pthread %08p\n", pthread->state, pthread);
+ }
+ pthread->state = PTH_RUNNABLE;
/* 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);
+ /* Again, GIANT WARNING: if you change this, change batch wakeup code */
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);
-}
-
-/* The calling thread is yielding. Do what you need to do to restart (like put
- * yourself on a runqueue), or do some accounting. Eventually, this might be a
- * little more generic than just yield. */
-void pth_thread_yield(struct uthread *uthread)
+ if (can_adjust_vcores)
+ vcore_request(threads_ready);
+}
+
+/* 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;
- 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);
+ /* 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);
- if (pthread->flags & PTHREAD_EXITING) {
- mcs_unlock_notifsafe(&queue_lock, &local_qn);
- uthread_destroy(uthread);
- } 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);
- }
-}
-
-void pth_thread_destroy(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;
-}
-
-void pth_preempt_pending(void)
-{
-}
-
-void pth_spawn_thread(uintptr_t pc_start, void *data)
-{
+ mcs_pdr_unlock(&queue_lock);
+ /* communicate to pth_thread_runnable */
+ pthread->state = PTH_BLK_PAUSED;
+ /* 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. */
+ pth_thread_runnable(uthread);
}
/* Restarts a uthread hanging off a syscall. For the simple pthread case, we
struct uthread *ut_restartee = (struct uthread*)sysc->u_data;
/* uthread stuff here: */
assert(ut_restartee);
- assert(ut_restartee->state == UT_BLOCKED);
- assert(ut_restartee->sysc == sysc);
+ assert(((struct pthread_tcb*)ut_restartee)->state == PTH_BLK_SYSC);
+ assert(ut_restartee->sysc == sysc); /* set in uthread.c */
ut_restartee->sysc = 0; /* so we don't 'reblock' on this later */
- uthread_runnable(ut_restartee);
+ pth_thread_runnable(ut_restartee);
}
/* This handler is usually run in vcore context, though I can imagine it being
{
struct syscall *sysc;
assert(in_vcore_context());
+ /* if we just got a bit (not a msg), it should be because the process is
+ * still an SCP and hasn't started using the MCP ev_q yet (using the simple
+ * ev_q and glibc's blockon) or because the bit is still set from an old
+ * ev_q (blocking syscalls from before we could enter vcore ctx). Either
+ * way, just return. Note that if you screwed up the pth ev_q and made it
+ * NO_MSG, you'll never notice (we used to assert(ev_msg)). */
+ if (!ev_msg)
+ return;
/* It's a bug if we don't have a msg (we're handling a syscall bit-event) */
assert(ev_msg);
/* Get the sysc from the message and just restart it */
* and is trying to block on sysc. Need to put it somewhere were we can wake it
* up when the sysc is done. For now, we'll have the kernel send us an event
* when the syscall is done. */
-void pth_blockon_sysc(struct syscall *sysc)
+void pth_thread_blockon_sysc(struct uthread *uthread, void *syscall)
{
+ struct syscall *sysc = (struct syscall*)syscall;
int old_flags;
bool need_to_restart = FALSE;
uint32_t vcoreid = vcore_id();
-
- 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);
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ pthread->state = PTH_BLK_SYSC;
+ 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;
+ sysc->u_data = uthread;
/* Register our vcore's syscall ev_q to hear about this syscall. */
if (!register_evq(sysc, sysc_mgmt[vcoreid].ev_q)) {
/* Lost the race with the call being done. The kernel won't send the
/* GIANT WARNING: do not touch the thread after this point. */
}
+void pth_thread_has_blocked(struct uthread *uthread, int flags)
+{
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ /* could imagine doing something with the flags. For now, we just treat all
+ * externally blocked reasons as 'MUTEX'. Whatever we do here, we are
+ * mostly communicating to our future selves in pth_thread_runnable(), which
+ * gets called by whoever triggered this callback */
+ pthread->state = PTH_BLK_MUTEX;
+ /* Just for yucks: */
+ if (flags == UTH_EXT_BLK_JUSTICE)
+ printf("For great justice!\n");
+}
+
+void pth_preempt_pending(void)
+{
+}
+
+void pth_spawn_thread(uintptr_t pc_start, void *data)
+{
+}
+
+/* Akaros pthread extensions / hacks */
+
+/* Tells the pthread 2LS to not change the number of vcores. This means it will
+ * neither request vcores nor yield vcores. Only used for testing. */
+void pthread_can_vcore_request(bool can)
+{
+ /* checked when we would request or yield */
+ can_adjust_vcores = can;
+}
+
/* Pthread interface stuff and helpers */
int pthread_attr_init(pthread_attr_t *a)
static void __pthread_free_stack(struct pthread_tcb *pt)
{
- assert(!munmap(pt->stacktop - pt->stacksize, pt->stacksize));
+ int ret = munmap(pt->stacktop - pt->stacksize, pt->stacksize);
+ assert(!ret);
}
static int __pthread_allocate_stack(struct pthread_tcb *pt)
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()) */
+void pthread_lib_init(void)
{
- struct pthread_tcb *pthread =
- (struct pthread_tcb*)uthread_create(__pthread_run, (void*)attr);
- if (!pthread)
- return -1;
+ uintptr_t mmap_block;
+ struct pthread_tcb *t;
+ int ret;
+ /* Some testing code might call this more than once (once for a slimmed down
+ * pth 2LS, and another from pthread_create(). Also, this is racy, but the
+ * first time through we are an SCP. */
+ init_once_racy(return);
+ assert(!in_multi_mode());
+ mcs_pdr_init(&queue_lock);
+ /* Create a pthread_tcb for the main thread */
+ ret = posix_memalign((void**)&t, __alignof__(struct pthread_tcb),
+ sizeof(struct pthread_tcb));
+ assert(!ret);
+ memset(t, 0, sizeof(struct pthread_tcb)); /* aggressively 0 for bugs */
+ t->id = get_next_pid();
+ t->stacksize = USTACK_NUM_PAGES * PGSIZE;
+ t->stacktop = (void*)USTACKTOP;
+ t->detached = TRUE;
+ t->state = PTH_RUNNING;
+ t->joiner = 0;
+ assert(t->id == 0);
+ /* Put the new pthread (thread0) on the active queue */
+ mcs_pdr_lock(&queue_lock);
+ 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() probably has
+ * happened, but don't rely on this. Careful if your 2LS somehow wants to
+ * have its init stuff use things like vcore stacks or TLSs, we'll need to
+ * change this. */
+ uthread_lib_init((struct uthread*)t);
+}
+
+int pthread_create(pthread_t *thread, const pthread_attr_t *attr,
+ void *(*start_routine)(void *), void *arg)
+{
+ run_once(pthread_lib_init());
+ /* Create the actual thread */
+ struct pthread_tcb *pthread;
+ int ret = posix_memalign((void**)&pthread, __alignof__(struct pthread_tcb),
+ sizeof(struct pthread_tcb));
+ assert(!ret);
+ memset(pthread, 0, sizeof(struct pthread_tcb)); /* aggressively 0 for bugs*/
+ pthread->stacksize = PTHREAD_STACK_SIZE; /* default */
+ pthread->state = PTH_CREATED;
+ pthread->id = get_next_pid();
+ pthread->detached = FALSE; /* default */
+ 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_ctx(&pthread->uthread.u_ctx, (uintptr_t)&__pthread_run,
+ (uintptr_t)(pthread->stacktop));
pthread->start_routine = start_routine;
pthread->arg = arg;
- uthread_runnable((struct uthread*)pthread);
+ /* Initialize the uthread */
+ uthread_init((struct uthread*)pthread);
+ pth_thread_runnable((struct uthread*)pthread);
*thread = pthread;
return 0;
}
-int pthread_join(pthread_t thread, void** retval)
+/* Helper that all pthread-controlled yield paths call. Just does some
+ * accounting. This is another example of how the much-loathed (and loved)
+ * active queue is keeping us honest. Need to export for sem and friends. */
+void __pthread_generic_yield(struct pthread_tcb *pthread)
+{
+ mcs_pdr_lock(&queue_lock);
+ threads_active--;
+ TAILQ_REMOVE(&active_queue, pthread, next);
+ mcs_pdr_unlock(&queue_lock);
+}
+
+/* Callback/bottom half of join, called from __uthread_yield (vcore context).
+ * join_target is who we are trying to join on (and who is calling exit). */
+static void __pth_join_cb(struct uthread *uthread, void *arg)
+{
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ struct pthread_tcb *join_target = (struct pthread_tcb*)arg;
+ struct pthread_tcb *temp_pth = 0;
+ __pthread_generic_yield(pthread);
+ /* We're trying to join, yield til we get woken up */
+ pthread->state = PTH_BLK_JOINING; /* could do this front-side */
+ /* Put ourselves in the join target's joiner slot. If we get anything back,
+ * we lost the race and need to wake ourselves. Syncs with __pth_exit_cb.*/
+ temp_pth = atomic_swap_ptr((void**)&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) { /* temp_pth != 0 means they exited first */
+ assert(temp_pth == 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(uthread); /* wake ourselves */
+ }
+}
+
+int pthread_join(struct pthread_tcb *join_target, void **retval)
{
/* 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
+ * join_target and he is already freed (which would have happened if he was
* detached. */
- if (thread->detached) {
+ if (join_target->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 (!join_target->joiner) {
+ uthread_yield(TRUE, __pth_join_cb, join_target);
+ /* When we return/restart, the thread will be done */
+ } else {
+ assert(join_target->joiner == join_target); /* sanity check */
+ }
if (retval)
- *retval = thread->retval;
- free(thread);
+ *retval = join_target->retval;
+ free(join_target);
return 0;
}
+/* Callback/bottom half of exit. Syncs with __pth_join_cb. Here's how it
+ * works: the slot for joiner is initially 0. Joiners try to swap themselves
+ * into that spot. Exiters try to put 'themselves' into it. Whoever gets 0
+ * back won the race. If the exiter lost the race, it must wake up the joiner
+ * (which was the value from temp_pth). If the joiner lost the race, it must
+ * wake itself up, and for sanity reasons can ensure the value from temp_pth is
+ * the join target). */
+static void __pth_exit_cb(struct uthread *uthread, void *junk)
+{
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ struct pthread_tcb *temp_pth = 0;
+ __pthread_generic_yield(pthread);
+ /* Catch some bugs */
+ pthread->state = PTH_EXITING;
+ /* Destroy the pthread */
+ uthread_cleanup(uthread);
+ /* Cleanup, mirroring pthread_create() */
+ __pthread_free_stack(pthread);
+ /* TODO: race on detach state (see join) */
+ 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);
+ }
+ }
+}
+
+void pthread_exit(void *ret)
+{
+ struct pthread_tcb *pthread = pthread_self();
+ pthread->retval = ret;
+ destroy_dtls();
+ uthread_yield(FALSE, __pth_exit_cb, 0);
+}
+
+/* Callback/bottom half of yield. For those writing these pth callbacks, the
+ * minimum is call generic, set state (communicate with runnable), then do
+ * something that causes it to be runnable in the future (or right now). */
+static void __pth_yield_cb(struct uthread *uthread, void *junk)
+{
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ __pthread_generic_yield(pthread);
+ pthread->state = PTH_BLK_YIELDING;
+ /* just immediately restart it */
+ pth_thread_runnable(uthread);
+}
+
+/* Cooperative yielding of the processor, to allow other threads to run */
int pthread_yield(void)
{
- uthread_yield(TRUE);
+ uthread_yield(TRUE, __pth_yield_cb, 0);
return 0;
}
return 0;
}
+/* Helper for spinning sync, returns TRUE if it is okay to keep spinning.
+ *
+ * Alternatives include:
+ * old_count <= num_vcores() (barrier code, pass in old_count as *state,
+ * but this only works if every awake pthread
+ * will belong to the barrier).
+ * just spin for a bit (use *state to track spins)
+ * FALSE (always is safe)
+ * etc...
+ * 'threads_ready' isn't too great since sometimes it'll be non-zero when it is
+ * about to become 0. We really want "I have no threads waiting to run that
+ * aren't going to run on their on unless this core yields instead of spins". */
+/* TODO: consider making this a 2LS op */
+static inline bool safe_to_spin(unsigned int *state)
+{
+ return !threads_ready;
+}
+
/* Set *spun to 0 when calling this the first time. It will yield after 'spins'
* calls. Use this for adaptive mutexes and such. */
static inline void spin_to_sleep(unsigned int spins, unsigned int *spun)
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;
int pthread_cond_init(pthread_cond_t *c, const pthread_condattr_t *a)
{
- c->attr = a;
- memset(c->waiters,0,sizeof(c->waiters));
- memset(c->in_use,0,sizeof(c->in_use));
- c->next_waiter = 0;
- return 0;
+ TAILQ_INIT(&c->waiters);
+ spin_pdr_init(&c->spdr_lock);
+ if (a) {
+ c->attr_pshared = a->pshared;
+ c->attr_clock = a->clock;
+ } else {
+ c->attr_pshared = PTHREAD_PROCESS_PRIVATE;
+ c->attr_clock = 0;
+ }
+ return 0;
}
int pthread_cond_destroy(pthread_cond_t *c)
{
- return 0;
+ return 0;
}
int pthread_cond_broadcast(pthread_cond_t *c)
{
- memset(c->waiters,0,sizeof(c->waiters));
- return 0;
+ unsigned int nr_woken = 0; /* assuming less than 4 bil threads */
+ struct pthread_queue restartees = TAILQ_HEAD_INITIALIZER(restartees);
+ struct pthread_tcb *pthread_i;
+ spin_pdr_lock(&c->spdr_lock);
+ /* moves all items from waiters onto the end of restartees */
+ TAILQ_CONCAT(&restartees, &c->waiters, next);
+ spin_pdr_unlock(&c->spdr_lock);
+ /* Do the work of pth_thread_runnable(). We're in uth context here, but I
+ * think it's okay. When we need to (when locking) we drop into VC ctx, as
+ * far as the kernel and other cores are concerned. */
+ TAILQ_FOREACH(pthread_i, &restartees, next) {
+ pthread_i->state = PTH_RUNNABLE;
+ nr_woken++;
+ }
+ /* Amortize the lock grabbing over all restartees */
+ mcs_pdr_lock(&queue_lock);
+ threads_ready += nr_woken;
+ TAILQ_CONCAT(&ready_queue, &restartees, next);
+ mcs_pdr_unlock(&queue_lock);
+ if (can_adjust_vcores)
+ vcore_request(threads_ready);
+ return 0;
}
+/* spec says this needs to work regardless of whether or not it holds the mutex
+ * already. */
int pthread_cond_signal(pthread_cond_t *c)
{
- int i;
- for(i = 0; i < MAX_PTHREADS; i++)
- {
- if(c->waiters[i])
- {
- c->waiters[i] = 0;
- break;
- }
- }
- return 0;
+ struct pthread_tcb *pthread;
+ spin_pdr_lock(&c->spdr_lock);
+ pthread = TAILQ_FIRST(&c->waiters);
+ if (!pthread) {
+ spin_pdr_unlock(&c->spdr_lock);
+ return 0;
+ }
+ TAILQ_REMOVE(&c->waiters, pthread, next);
+ spin_pdr_unlock(&c->spdr_lock);
+ pth_thread_runnable((struct uthread*)pthread);
+ return 0;
}
-int pthread_cond_wait(pthread_cond_t *c, pthread_mutex_t *m)
+/* Communicate btw cond_wait and its callback */
+struct cond_junk {
+ pthread_cond_t *c;
+ pthread_mutex_t *m;
+};
+
+/* Callback/bottom half of cond wait. For those writing these pth callbacks,
+ * the minimum is call generic, set state (communicate with runnable), then do
+ * something that causes it to be runnable in the future (or right now). */
+static void __pth_wait_cb(struct uthread *uthread, void *junk)
{
- uint32_t old_waiter = c->next_waiter;
- uint32_t my_waiter = c->next_waiter;
-
- //allocate a slot
- while (atomic_swap_u32(& (c->in_use[my_waiter]), SLOT_IN_USE) == SLOT_IN_USE)
- {
- my_waiter = (my_waiter + 1) % MAX_PTHREADS;
- assert (old_waiter != my_waiter); // do not want to wrap around
- }
- c->waiters[my_waiter] = WAITER_WAITING;
- c->next_waiter = (my_waiter+1) % MAX_PTHREADS; // race on next_waiter but ok, because it is advisary
-
- pthread_mutex_unlock(m);
-
- volatile int* poll = &c->waiters[my_waiter];
- while(*poll);
- c->in_use[my_waiter] = SLOT_FREE;
- pthread_mutex_lock(m);
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ pthread_cond_t *c = ((struct cond_junk*)junk)->c;
+ pthread_mutex_t *m = ((struct cond_junk*)junk)->m;
+ /* this removes us from the active list; we can reuse next below */
+ __pthread_generic_yield(pthread);
+ pthread->state = PTH_BLK_MUTEX;
+ spin_pdr_lock(&c->spdr_lock);
+ TAILQ_INSERT_TAIL(&c->waiters, pthread, next);
+ spin_pdr_unlock(&c->spdr_lock);
+ pthread_mutex_unlock(m);
+}
- return 0;
+int pthread_cond_wait(pthread_cond_t *c, pthread_mutex_t *m)
+{
+ struct cond_junk local_junk;
+ local_junk.c = c;
+ local_junk.m = m;
+ uthread_yield(TRUE, __pth_wait_cb, &local_junk);
+ pthread_mutex_lock(m);
+ return 0;
}
int pthread_condattr_init(pthread_condattr_t *a)
{
- a = PTHREAD_PROCESS_PRIVATE;
- return 0;
+ a->pshared = PTHREAD_PROCESS_PRIVATE;
+ a->clock = 0;
+ return 0;
}
int pthread_condattr_destroy(pthread_condattr_t *a)
{
- return 0;
+ return 0;
+}
+
+int pthread_condattr_getpshared(pthread_condattr_t *a, int *s)
+{
+ *s = a->pshared;
+ return 0;
}
int pthread_condattr_setpshared(pthread_condattr_t *a, int s)
{
- a->pshared = s;
- return 0;
+ a->pshared = s;
+ if (s == PTHREAD_PROCESS_SHARED) {
+ printf("Warning: we don't do shared pthread condvars btw diff MCPs\n");
+ return -1;
+ }
+ return 0;
}
-int pthread_condattr_getpshared(pthread_condattr_t *a, int *s)
+int pthread_condattr_getclock(const pthread_condattr_t *attr,
+ clockid_t *clock_id)
{
- *s = a->pshared;
- return 0;
+ *clock_id = attr->clock;
+}
+
+int pthread_condattr_setclock(pthread_condattr_t *attr, clockid_t clock_id)
+{
+ printf("Warning: we don't do pthread condvar clock stuff\n");
+ attr->clock = clock_id;
}
pthread_t pthread_self()
return t1 == t2;
}
-/* This function cannot be migrated to a different vcore by the userspace
- * scheduler. Will need to sort that shit out. */
-void pthread_exit(void *ret)
-{
- struct pthread_tcb *pthread = pthread_self();
- pthread->retval = ret;
- /* So our pth_thread_yield knows we want to exit */
- pthread->flags |= PTHREAD_EXITING;
- uthread_yield(FALSE);
-}
-
int pthread_once(pthread_once_t* once_control, void (*init_routine)(void))
{
if (atomic_swap_u32(once_control, 1) == 0)
return 0;
}
-int pthread_barrier_init(pthread_barrier_t* b, const pthread_barrierattr_t* a, int count)
+int pthread_barrier_init(pthread_barrier_t *b,
+ const pthread_barrierattr_t *a, int count)
{
- b->nprocs = b->count = count;
- b->sense = 0;
- pthread_mutex_init(&b->pmutex, 0);
- return 0;
+ b->total_threads = count;
+ b->sense = 0;
+ atomic_set(&b->count, count);
+ spin_pdr_init(&b->lock);
+ TAILQ_INIT(&b->waiters);
+ b->nr_waiters = 0;
+ return 0;
}
-int pthread_barrier_wait(pthread_barrier_t* b)
-{
- unsigned int spinner = 0;
- int ls = !b->sense;
-
- pthread_mutex_lock(&b->pmutex);
- int count = --b->count;
- pthread_mutex_unlock(&b->pmutex);
+struct barrier_junk {
+ pthread_barrier_t *b;
+ int ls;
+};
- if(count == 0)
- {
- printd("Thread %d is last to hit the barrier, resetting...\n", pthread_self()->id);
- b->count = b->nprocs;
- wmb();
- b->sense = ls;
- return PTHREAD_BARRIER_SERIAL_THREAD;
- }
- else
- {
- while(b->sense != ls) {
- cpu_relax();
- spin_to_sleep(PTHREAD_BARRIER_SPINS, &spinner);
- }
- return 0;
- }
+/* Callback/bottom half of barrier. */
+static void __pth_barrier_cb(struct uthread *uthread, void *junk)
+{
+ struct pthread_tcb *pthread = (struct pthread_tcb*)uthread;
+ pthread_barrier_t *b = ((struct barrier_junk*)junk)->b;
+ int ls = ((struct barrier_junk*)junk)->ls;
+ /* Removes from active list, we can reuse. must also restart */
+ __pthread_generic_yield(pthread);
+ /* TODO: if we used a trylock, we could bail as soon as we see sense */
+ spin_pdr_lock(&b->lock);
+ /* If sense is ls (our free value), we lost the race and shouldn't sleep */
+ if (b->sense == ls) {
+ /* TODO: i'd like to fast-path the wakeup, skipping pth_runnable */
+ pthread->state = PTH_BLK_YIELDING; /* not sure which state for this */
+ spin_pdr_unlock(&b->lock);
+ pth_thread_runnable(uthread);
+ return;
+ }
+ /* otherwise, we sleep */
+ pthread->state = PTH_BLK_MUTEX; /* TODO: consider ignoring this */
+ TAILQ_INSERT_TAIL(&b->waiters, pthread, next);
+ b->nr_waiters++;
+ spin_pdr_unlock(&b->lock);
+}
+
+/* We assume that the same threads participating in the barrier this time will
+ * also participate next time. Imagine a thread stopped right after its fetch
+ * and add - we know it is coming through eventually. We finish and change the
+ * sense, which should allow the delayed thread to eventually break through.
+ * But if another n threads come in first, we'll set the sense back to the old
+ * value, thereby catching the delayed thread til the next barrier.
+ *
+ * A note on preemption: if any thread gets preempted and it is never dealt
+ * with, eventually we deadlock, with all threads waiting on the last one to
+ * enter (and any stragglers from one run will be the last in the next run).
+ * One way or another, we need to handle preemptions. The current 2LS requests
+ * an IPI for a preempt, so we'll be fine. Any other strategies will need to
+ * consider how barriers work. Any time we sleep, we'll be okay (since that
+ * frees up our core to handle preemptions/run other threads. */
+int pthread_barrier_wait(pthread_barrier_t *b)
+{
+ unsigned int spin_state = 0;
+ int ls = !b->sense; /* when b->sense is the value we read, then we're free*/
+ int nr_waiters;
+ struct pthread_queue restartees = TAILQ_HEAD_INITIALIZER(restartees);
+ struct pthread_tcb *pthread_i;
+ struct barrier_junk local_junk;
+
+ long old_count = atomic_fetch_and_add(&b->count, -1);
+
+ if (old_count == 1) {
+ printd("Thread %d is last to hit the barrier, resetting...\n",
+ pthread_self()->id);
+ /* TODO: we might want to grab the lock right away, so a few short
+ * circuit faster? */
+ atomic_set(&b->count, b->total_threads);
+ /* we still need to maintain ordering btw count and sense, in case
+ * another thread doesn't sleep (if we wrote sense first, they could
+ * break out, race around, and muck with count before it is time) */
+ /* wmb(); handled by the spin lock */
+ spin_pdr_lock(&b->lock);
+ /* Sense is only protected in addition to decisions to sleep */
+ b->sense = ls; /* set to free everyone */
+ /* All access to nr_waiters is protected by the lock */
+ if (!b->nr_waiters) {
+ spin_pdr_unlock(&b->lock);
+ return PTHREAD_BARRIER_SERIAL_THREAD;
+ }
+ TAILQ_CONCAT(&restartees, &b->waiters, next);
+ nr_waiters = b->nr_waiters;
+ b->nr_waiters = 0;
+ spin_pdr_unlock(&b->lock);
+ /* TODO: do we really need this state tracking? */
+ TAILQ_FOREACH(pthread_i, &restartees, next)
+ pthread_i->state = PTH_RUNNABLE;
+ /* bulk restart waiters (skipping pth_thread_runnable()) */
+ mcs_pdr_lock(&queue_lock);
+ threads_ready += nr_waiters;
+ TAILQ_CONCAT(&ready_queue, &restartees, next);
+ mcs_pdr_unlock(&queue_lock);
+ if (can_adjust_vcores)
+ vcore_request(threads_ready);
+ return PTHREAD_BARRIER_SERIAL_THREAD;
+ } else {
+ /* Spin if there are no other threads to run. No sense sleeping */
+ do {
+ if (b->sense == ls)
+ return 0;
+ cpu_relax();
+ } while (safe_to_spin(&spin_state));
+
+ /* Try to sleep, when we wake/return, we're free to go */
+ local_junk.b = b;
+ local_junk.ls = ls;
+ uthread_yield(TRUE, __pth_barrier_cb, &local_junk);
+ // assert(b->sense == ls);
+ return 0;
+ }
}
-int pthread_barrier_destroy(pthread_barrier_t* b)
+int pthread_barrier_destroy(pthread_barrier_t *b)
{
- pthread_mutex_destroy(&b->pmutex);
- return 0;
+ assert(TAILQ_EMPTY(&b->waiters));
+ assert(!b->nr_waiters);
+ /* Free any locks (if we end up using an MCS) */
+ return 0;
}
int pthread_detach(pthread_t thread)
thread->detached = TRUE;
return 0;
}
+
+int pthread_kill (pthread_t __threadid, int __signo)
+{
+ printf("pthread_kill is not yet implemented!");
+ return -1;
+}
+
+
+int pthread_sigmask(int how, const sigset_t *set, sigset_t *oset)
+{
+ printf("pthread_sigmask is not yet implemented!");
+ return -1;
+}
+
+int pthread_sigqueue(pthread_t *thread, int sig, const union sigval value)
+{
+ printf("pthread_sigqueue is not yet implemented!");
+ return -1;
+}
+
+int pthread_key_create(pthread_key_t *key, void (*destructor)(void*))
+{
+ *key = dtls_key_create(destructor);
+ assert(key);
+ return 0;
+}
+
+int pthread_key_delete(pthread_key_t key)
+{
+ dtls_key_delete(key);
+ return 0;
+}
+
+void *pthread_getspecific(pthread_key_t key)
+{
+ return get_dtls(key);
+}
+
+int pthread_setspecific(pthread_key_t key, const void *value)
+{
+ set_dtls(key, (void*)value);
+ return 0;
+}
+
projects / akaros.git / blobdiff