path: root/litmus/locking.c

#include <linux/sched.h>
#include <litmus/litmus.h>
#include <litmus/fdso.h>

#ifdef CONFIG_LITMUS_LOCKING

#include <linux/sched.h>
#include <litmus/litmus.h>
#include <litmus/sched_plugin.h>
#include <litmus/trace.h>
#include <litmus/litmus.h>
#include <litmus/wait.h>

#ifdef CONFIG_LITMUS_DGL_SUPPORT
#include <linux/uaccess.h>
#endif

#if defined(CONFIG_LITMUS_AFFINITY_LOCKING) && defined(CONFIG_LITMUS_NVIDIA)
#include <litmus/gpu_affinity.h>
#endif

static int create_generic_lock(void** obj_ref, obj_type_t type, void* __user arg);
static int open_generic_lock(struct od_table_entry* entry, void* __user arg);
static int close_generic_lock(struct od_table_entry* entry);
static void destroy_generic_lock(obj_type_t type, void* sem);

struct fdso_ops generic_lock_ops = {
	.create  = create_generic_lock,
	.open    = open_generic_lock,
	.close   = close_generic_lock,
	.destroy = destroy_generic_lock,
};

static atomic_t lock_id_gen = ATOMIC_INIT(0);


static inline bool is_lock(struct od_table_entry* entry)
{
	return entry->class == &generic_lock_ops;
}

static inline struct litmus_lock* get_lock(struct od_table_entry* entry)
{
	BUG_ON(!is_lock(entry));
	return (struct litmus_lock*) entry->obj->obj;
}

static  int create_generic_lock(void** obj_ref, obj_type_t type, void* __user arg)
{
	struct litmus_lock* lock;
	int err;

	err = litmus->allocate_lock(&lock, type, arg);
	if (err == 0) {
#ifdef CONFIG_LITMUS_NESTED_LOCKING
		lock->nest.lock = lock;
		lock->nest.hp_waiter_eff_prio = NULL;

		INIT_BINHEAP_NODE(&lock->nest.hp_binheap_node);
		if(!lock->nest.hp_waiter_ptr) {
			TRACE_CUR("BEWARE: hp_waiter_ptr should probably not be NULL in "
					  "most cases. (exception: IKGLP donors)\n");
		}
#endif
		lock->type = type;
		lock->ident = atomic_inc_return(&lock_id_gen);
		*obj_ref = lock;

		TRACE_CUR("Lock %d (%p) created. Type = %d\n.", lock->ident, lock, type);

		if (lock->proc && lock->proc->add) {
			lock->proc->add(lock);
		}
    }
	return err;
}

static int open_generic_lock(struct od_table_entry* entry, void* __user arg)
{
	struct litmus_lock* lock = get_lock(entry);
	if (lock->ops->open)
		return lock->ops->open(lock, arg);
	else
		return 0; /* default: any task can open it */
}

static int close_generic_lock(struct od_table_entry* entry)
{
	struct litmus_lock* lock = get_lock(entry);
	if (lock->ops->close) {

		if (lock->proc && lock->proc->remove) {
			lock->proc->remove(lock);
		}

		return lock->ops->close(lock);
	}
	else
		return 0; /* default: closing succeeds */
}

static void destroy_generic_lock(obj_type_t type, void* obj)
{
	struct litmus_lock* lock = (struct litmus_lock*) obj;
	lock->ops->deallocate(lock);
}

asmlinkage long sys_litmus_lock(int lock_od)
{
	long err = -EINVAL;
	struct od_table_entry* entry;
	struct litmus_lock* l;

	TS_SYSCALL_IN_START;

	TS_SYSCALL_IN_END;

	TS_LOCK_START;

	entry = get_entry_for_od(lock_od);
	if (entry && is_lock(entry)) {
		l = get_lock(entry);
		//TRACE_CUR("attempts to lock 0x%p\n", l);
		TRACE_CUR("Attempts to lock %d\n", l->ident);
		err = l->ops->lock(l);
		if (!err) {
			TRACE_CUR("Got lock %d\n", l->ident);
		}
	}

	/* Note: task my have been suspended or preempted in between!  Take
	 * this into account when computing overheads. */
	TS_LOCK_END;

	TS_SYSCALL_OUT_START;

	return err;
}

asmlinkage long sys_litmus_unlock(int lock_od)
{
	long err = -EINVAL;
	struct od_table_entry* entry;
	struct litmus_lock* l;

	TS_SYSCALL_IN_START;

	TS_SYSCALL_IN_END;

	TS_UNLOCK_START;

	entry = get_entry_for_od(lock_od);
	if (entry && is_lock(entry)) {
		l = get_lock(entry);
		//TRACE_CUR("attempts to unlock 0x%p\n", l);
		TRACE_CUR("Attempts to unlock %d\n", l->ident);
		err = l->ops->unlock(l);
		if (!err) {
			TRACE_CUR("Unlocked %d\n", l->ident);
		}
	}

	/* Note: task my have been preempted in between!  Take this into
	 * account when computing overheads. */
	TS_UNLOCK_END;

	TS_SYSCALL_OUT_START;

	return err;
}

struct task_struct* __waitqueue_remove_first(wait_queue_head_t *wq)
{
	wait_queue_t* q;
	struct task_struct* t = NULL;

	if (waitqueue_active(wq)) {
		q = list_entry(wq->task_list.next,
			       wait_queue_t, task_list);
		t = (struct task_struct*) q->private;
		__remove_wait_queue(wq, q);
	}
	return(t);
}

#ifdef CONFIG_LITMUS_NESTED_LOCKING

void print_hp_waiters(struct binheap_node* n, int depth)
{
	struct litmus_lock *l;
	struct nested_info *nest;
	char padding[81] = "                                                                                ";
	struct task_struct *hp = NULL;
	struct task_struct *hp_eff = NULL;
	struct task_struct *node_prio = NULL;


	if(n == NULL) {
		TRACE("+-> %p\n", NULL);
		return;
	}

	nest = binheap_entry(n, struct nested_info, hp_binheap_node);
	l = nest->lock;

	if(depth*2 <= 80)
		padding[depth*2] = '\0';

	if(nest->hp_waiter_ptr && *(nest->hp_waiter_ptr)) {
		hp = *(nest->hp_waiter_ptr);

		if(tsk_rt(hp)->inh_task) {
			hp_eff = tsk_rt(hp)->inh_task;
		}
	}

	node_prio = nest->hp_waiter_eff_prio;

	TRACE("%s+-> %s/%d [waiter = %s/%d] [waiter's inh = %s/%d] (lock = %d)\n",
		  padding,
		  (node_prio) ? node_prio->comm : "null",
		  (node_prio) ? node_prio->pid : 0,
		  (hp) ? hp->comm : "null",
		  (hp) ? hp->pid : 0,
		  (hp_eff) ? hp_eff->comm : "null",
		  (hp_eff) ? hp_eff->pid : 0,
		  l->ident);

    if(n->left) print_hp_waiters(n->left, depth+1);
    if(n->right) print_hp_waiters(n->right, depth+1);
}
#endif


#ifdef CONFIG_LITMUS_DGL_SUPPORT

void select_next_lock(dgl_wait_state_t* dgl_wait /*, struct litmus_lock* prev_lock*/)
{
	/*
	 We pick the next lock in reverse order. This causes inheritance propagation
	 from locks received earlier to flow in the same direction as regular nested
	 locking. This might make fine-grain DGL easier in the future.
	 */

	BUG_ON(tsk_rt(dgl_wait->task)->blocked_lock);

	//WARN_ON(dgl_wait->locks[dgl_wait->last_primary] != prev_lock);

	// note reverse order
	for(dgl_wait->last_primary = dgl_wait->last_primary - 1;
		dgl_wait->last_primary >= 0;
		--(dgl_wait->last_primary)){
		if(!dgl_wait->locks[dgl_wait->last_primary]->ops->is_owner(
				dgl_wait->locks[dgl_wait->last_primary], dgl_wait->task)) {

			tsk_rt(dgl_wait->task)->blocked_lock =
					dgl_wait->locks[dgl_wait->last_primary];
			mb();

			TRACE_CUR("New blocked lock is %d\n",
					  dgl_wait->locks[dgl_wait->last_primary]->ident);

			break;
		}
	}
}

int dgl_wake_up(wait_queue_t *wq_node, unsigned mode, int sync, void *key)
{
	// should never be called.
	BUG();
	return 1;
}

void __waitqueue_dgl_remove_first(wait_queue_head_t *wq,
								  dgl_wait_state_t** dgl_wait,
								  struct task_struct **task)
{
	wait_queue_t *q;

	*dgl_wait = NULL;
	*task = NULL;

	if (waitqueue_active(wq)) {
		q = list_entry(wq->task_list.next,
					   wait_queue_t, task_list);

		if(q->func == dgl_wake_up) {
			*dgl_wait = (dgl_wait_state_t*) q->private;
		}
		else {
			*task = (struct task_struct*) q->private;
		}

		__remove_wait_queue(wq, q);
	}
}

void init_dgl_waitqueue_entry(wait_queue_t *wq_node, dgl_wait_state_t* dgl_wait)
{
	init_waitqueue_entry(wq_node, dgl_wait->task);
	wq_node->private = dgl_wait;
	wq_node->func = dgl_wake_up;
}

#ifdef CONFIG_SCHED_DEBUG_TRACE
static void snprintf_dgl(char* buf, size_t bsz, struct litmus_lock* dgl_locks[], int sz)
{
	int i;
	char* ptr;

	ptr = buf;
	for(i = 0; i < sz && ptr < buf+bsz; ++i)
	{
		struct litmus_lock *l = dgl_locks[i];
		int remaining = bsz - (ptr-buf);
		int written;

		if(i == 0)
			written = snprintf(ptr, remaining, "%d ", l->ident);
		else if(i == sz - 1)
			written = snprintf(ptr, remaining, " %d", l->ident);
		else
			written = snprintf(ptr, remaining, " %d ", l->ident);
		ptr += written;
	}
}
#endif

static long do_litmus_dgl_lock(dgl_wait_state_t *dgl_wait)
{
	int i;
	unsigned long irqflags; //, dummyflags;
	raw_spinlock_t *dgl_lock;

#ifdef CONFIG_SCHED_DEBUG_TRACE
	char dglstr[CONFIG_LITMUS_MAX_DGL_SIZE*5];
	snprintf_dgl(dglstr, sizeof(dglstr), dgl_wait->locks, dgl_wait->size);
	TRACE_CUR("Locking DGL with size %d: %s\n", dgl_wait->size, dglstr);
#endif

	dgl_lock = litmus->get_dgl_spinlock(dgl_wait->task);

	BUG_ON(dgl_wait->task != current);

	raw_spin_lock_irqsave(dgl_lock, irqflags);

	dgl_wait->nr_remaining = dgl_wait->size;

	// try to acquire each lock.  enqueue (non-blocking) if it is unavailable.
	for(i = 0; i < dgl_wait->size; ++i) {
		struct litmus_lock *l = dgl_wait->locks[i];

		// dgl_lock() must set task state to TASK_UNINTERRUPTIBLE if task blocks.

		if(l->ops->dgl_lock(l, dgl_wait, &dgl_wait->wq_nodes[i])) {
			--(dgl_wait->nr_remaining);
			TRACE_CUR("Acquired lock %d immediatly.\n", l->ident);
		}
	}

	if(dgl_wait->nr_remaining == 0) {
		// acquired entire group immediatly
		TRACE_CUR("Acquired all locks in DGL immediatly!\n");
	}
	else {

		TRACE_CUR("As many as %d locks in DGL are pending. Suspending.\n",
				  dgl_wait->nr_remaining);

#if defined(CONFIG_LITMUS_AFFINITY_LOCKING) && defined(CONFIG_LITMUS_NVIDIA)
		// KLUDGE: don't count this suspension as time in the critical gpu
		// critical section
		if(tsk_rt(dgl_wait->task)->held_gpus) {
			tsk_rt(dgl_wait->task)->suspend_gpu_tracker_on_block = 1;
		}
#endif

		// note reverse order.  see comments in select_next_lock for reason.
		for(i = dgl_wait->size - 1; i >= 0; --i) {
			struct litmus_lock *l = dgl_wait->locks[i];
			if(!l->ops->is_owner(l, dgl_wait->task)) {  // double-check to be thread safe

				TRACE_CUR("Activating priority inheritance on lock %d\n",
						  l->ident);

				TS_DGL_LOCK_SUSPEND;

				l->ops->enable_priority(l, dgl_wait);
				dgl_wait->last_primary = i;

				TRACE_CUR("Suspending for lock %d\n", l->ident);

				raw_spin_unlock_irqrestore(dgl_lock, irqflags);  // free dgl_lock before suspending

				suspend_for_lock(); // suspend!!!
				//schedule();  // suspend!!!

				TS_DGL_LOCK_RESUME;

				TRACE_CUR("Woken up from DGL suspension.\n");

				goto all_acquired;  // we should hold all locks when we wake up.
			}
		}

		TRACE_CUR("Didn't have to suspend after all, but calling schedule() anyway.\n");
		//BUG();
	}

	raw_spin_unlock_irqrestore(dgl_lock, irqflags);

all_acquired:

	// FOR SANITY CHECK FOR TESTING
//	for(i = 0; i < dgl_wait->size; ++i) {
//		struct litmus_lock *l = dgl_wait->locks[i];
//		BUG_ON(!l->ops->is_owner(l, dgl_wait->task));
//	}

	TRACE_CUR("Acquired entire DGL\n");

	return 0;
}

static int supports_dgl(struct litmus_lock *l)
{
	struct litmus_lock_ops* ops = l->ops;

	return (ops->dgl_lock			&&
			ops->is_owner			&&
			ops->enable_priority);
}

asmlinkage long sys_litmus_dgl_lock(void* __user usr_dgl_ods, int dgl_size)
{
	struct task_struct *t = current;
	long err = -EINVAL;
	int dgl_ods[MAX_DGL_SIZE];
	int i;

	dgl_wait_state_t dgl_wait_state;  // lives on the stack until all resources in DGL are held.

	if(dgl_size > MAX_DGL_SIZE || dgl_size < 1)
		goto out;

	if(!access_ok(VERIFY_READ, usr_dgl_ods, dgl_size*(sizeof(int))))
		goto out;

	if(__copy_from_user(&dgl_ods, usr_dgl_ods, dgl_size*(sizeof(int))))
		goto out;

	if (!is_realtime(t)) {
		err = -EPERM;
		goto out;
	}

	for(i = 0; i < dgl_size; ++i) {
		struct od_table_entry *entry = get_entry_for_od(dgl_ods[i]);
		if(entry && is_lock(entry)) {
			dgl_wait_state.locks[i] = get_lock(entry);
			if(!supports_dgl(dgl_wait_state.locks[i])) {
				TRACE_CUR("Lock %d does not support all required DGL operations.\n",
						  dgl_wait_state.locks[i]->ident);
				goto out;
			}
		}
		else {
			TRACE_CUR("Invalid lock identifier\n");
			goto out;
		}
	}

	dgl_wait_state.task = t;
	dgl_wait_state.size = dgl_size;

	TS_DGL_LOCK_START;
	err = do_litmus_dgl_lock(&dgl_wait_state);

	/* Note: task my have been suspended or preempted in between!  Take
	 * this into account when computing overheads. */
	TS_DGL_LOCK_END;

out:
	return err;
}

static long do_litmus_dgl_unlock(struct litmus_lock* dgl_locks[], int dgl_size)
{
	int i;
	long err = 0;

#ifdef CONFIG_SCHED_DEBUG_TRACE
	char dglstr[CONFIG_LITMUS_MAX_DGL_SIZE*5];
	snprintf_dgl(dglstr, sizeof(dglstr), dgl_locks, dgl_size);
	TRACE_CUR("Unlocking a DGL with size %d: %s\n", dgl_size, dglstr);
#endif

	for(i = dgl_size - 1; i >= 0; --i) {  // unlock in reverse order

		struct litmus_lock *l = dgl_locks[i];
		long tmp_err;

		TRACE_CUR("Unlocking lock %d of DGL.\n", l->ident);

		tmp_err = l->ops->unlock(l);

		if(tmp_err) {
			TRACE_CUR("There was an error unlocking %d: %d.\n", l->ident, tmp_err);
			err = tmp_err;
		}
	}

	TRACE_CUR("DGL unlocked. err = %d\n", err);

	return err;
}

asmlinkage long sys_litmus_dgl_unlock(void* __user usr_dgl_ods, int dgl_size)
{
	long err = -EINVAL;
	int dgl_ods[MAX_DGL_SIZE];
	struct od_table_entry* entry;
	int i;

	struct litmus_lock* dgl_locks[MAX_DGL_SIZE];

	if(dgl_size > MAX_DGL_SIZE || dgl_size < 1)
		goto out;

	if(!access_ok(VERIFY_READ, usr_dgl_ods, dgl_size*(sizeof(int))))
		goto out;

	if(__copy_from_user(&dgl_ods, usr_dgl_ods, dgl_size*(sizeof(int))))
		goto out;

	for(i = 0; i < dgl_size; ++i) {
		entry = get_entry_for_od(dgl_ods[i]);
		if(entry && is_lock(entry)) {
			dgl_locks[i] = get_lock(entry);
			if(!supports_dgl(dgl_locks[i])) {
				TRACE_CUR("Lock %d does not support all required DGL operations.\n",
						  dgl_locks[i]->ident);
				goto out;
			}
		}
		else {
			TRACE_CUR("Invalid lock identifier\n");
			goto out;
		}
	}

	TS_DGL_UNLOCK_START;
	err = do_litmus_dgl_unlock(dgl_locks, dgl_size);

	/* Note: task my have been suspended or preempted in between!  Take
	 * this into account when computing overheads. */
	TS_DGL_UNLOCK_END;

out:
	return err;
}

#else  // CONFIG_LITMUS_DGL_SUPPORT

asmlinkage long sys_litmus_dgl_lock(void* __user usr_dgl_ods, int dgl_size)
{
	return -ENOSYS;
}

asmlinkage long sys_litmus_dgl_unlock(void* __user usr_dgl_ods, int dgl_size)
{
	return -ENOSYS;
}

#endif

unsigned int __add_wait_queue_prio_exclusive(
	wait_queue_head_t* head,
	prio_wait_queue_t *new)
{
	struct list_head *pos;
	unsigned int passed = 0;

	new->wq.flags |= WQ_FLAG_EXCLUSIVE;

	/* find a spot where the new entry is less than the next */
	list_for_each(pos, &head->task_list) {
		prio_wait_queue_t* queued = list_entry(pos, prio_wait_queue_t,
						       wq.task_list);

		if (unlikely(lt_before(new->priority, queued->priority) ||
			     (new->priority == queued->priority &&
			      new->tie_breaker < queued->tie_breaker))) {
			/* pos is not less than new, thus insert here */
			__list_add(&new->wq.task_list, pos->prev, pos);
			goto out;
		}
		passed++;
	}

	/* if we get to this point either the list is empty or every entry
	 * queued element is less than new.
	 * Let's add new to the end. */
	list_add_tail(&new->wq.task_list, &head->task_list);
out:
	return passed;
}


void suspend_for_lock(void)
{
#if defined(CONFIG_REALTIME_AUX_TASKS) || defined(CONFIG_LITMUS_NVIDIA)
	struct task_struct *t = current;
#endif

#ifdef CONFIG_REALTIME_AUX_TASKS
	unsigned int aux_restore = 0;
	unsigned int aux_hide;
#endif

#ifdef CONFIG_LITMUS_NVIDIA
	unsigned int gpu_restore = 0;
	unsigned int gpu_hide;
#endif

#ifdef CONFIG_REALTIME_AUX_TASKS
	if (tsk_rt(t)->has_aux_tasks) {
		/* hide from aux tasks so they can't inherit our priority when we block
		 * for a litmus lock. inheritance is already going to a litmus lock
		 * holder. */
		aux_hide = tsk_rt(t)->hide_from_aux_tasks;
		aux_restore = 1;
		tsk_rt(t)->hide_from_aux_tasks = 1;
	}
#endif

#ifdef CONFIG_LITMUS_NVIDIA
	if (tsk_rt(t)->held_gpus) {
		gpu_hide = tsk_rt(t)->hide_from_gpu;
		gpu_restore = 1;
		tsk_rt(t)->hide_from_gpu = 1;
	}
#endif

	schedule();

#ifdef CONFIG_LITMUS_NVIDIA
	if (gpu_restore) {
		/* restore our state */
		tsk_rt(t)->hide_from_gpu = gpu_hide;
	}
#endif

#ifdef CONFIG_REALTIME_AUX_TASKS
	if (aux_restore) {
		/* restore our state */
		tsk_rt(t)->hide_from_aux_tasks = aux_hide;
	}
#endif
}


#else  // CONFIG_LITMUS_LOCKING

struct fdso_ops generic_lock_ops = {};

asmlinkage long sys_litmus_lock(int sem_od)
{
	return -ENOSYS;
}

asmlinkage long sys_litmus_unlock(int sem_od)
{
	return -ENOSYS;
}

#endif