/* SPDX-License-Identifier: GPL-2.0 */

/** \file sched.c Paraslash's scheduling functions. */

#include "para.h"
#include "ipc.h"
#include "fd.h"
#include "list.h"
#include "sched.h"
#include "string.h"
#include "time.h"
#include "error.h"

/**
 * The possible states of a task.
 *
 * In addition to the states listed here, a task may also enter zombie state.
 * This happens when its ->post_monitor function returns negative, the ->status
 * field is then set to this return value. Such tasks are not scheduled any
 * more (i.e. ->pre_monitor() and ->post_monitor() are no longer called), but
 * they stay on the scheduler task list until \ref task_reap() or
 * \ref sched_shutdown() is called.
 */
enum task_status {
	/** Task has been reaped and may be removed from the task list. */
	TS_DEAD,
	/** Task is active. */
	TS_RUNNING,
};

struct task {
	/** A copy of the task name supplied when the task was registered. */
	char *name;
	/** Copied during task_register(). */
	struct task_info info;
	/* TS_RUNNING, TS_DEAD, or zombie (negative value). */
	int status;
	/** Position of the task in the task list of the scheduler. */
	struct list_head node;
	/** If less than zero, the task was notified by another task. */
	int notification;
};

static struct timeval now_struct;
const struct timeval *now = &now_struct;

/* Internal representation of a paraslash scheduler instance. */
struct sched {
	/* Initial value (in milliseconds) before any pre_monitor call. */
	int default_timeout;
	/* The timeout (also in milliseconds) for the next iteration. */
	int timeout;
	/* Passed to poll(2). */
	struct pollfd *pfd;
	/* Number of elements in the above array, passed to poll(2). */
	unsigned pfd_array_len;
	/* Number of fds registered for monitoring so far. */
	unsigned num_pfds;
	/* Maps fds to indices of the pfd array. */
	unsigned *pidx;
	/* Number of elements in the above pidx array. */
	unsigned pidx_array_len;
	/* Either the application-supplied function or xpoll() of fd.c. */
	int (*poll_function)(struct pollfd *fds, nfds_t nfds, int timeout);
	/* Tasks which have been registered to the scheduler. */
	struct list_head task_list;
};

/**
 * Allocate and initialize a scheduler instance.
 *
 * \param poll_function Optional.
 *
 * If NULL is passed as the poll function pointer, the \ref xpoll() wrapper
 * is used to wait for events on the file descriptors to be monitored.
 *
 * \return A pointer to the new instance that can be used to register tasks
 * or to start scheduling. The only possible error is allocation failure,
 * in which case the function aborts. Thus, it never returns NULL.
 */
struct sched *sched_new(int (*poll_function)(struct pollfd *, nfds_t, int))
{
	struct sched *s = zalloc(sizeof(*s));
	s->default_timeout = 1000;
	init_list_head(&s->task_list);
	s->poll_function = poll_function? poll_function : xpoll;
	return s;
}

static void sched_pre_monitor(struct sched *s)
{
	struct task *t, *tmp;

	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
		if (t->status < 0)
			continue;
		if (t->notification != 0)
			sched_min_delay(s);
		if (t->info.pre_monitor)
			t->info.pre_monitor(s, t->info.context);
	}
}

static void unlink_and_free_task(struct task *t)
{
	list_del(&t->node);
	free(t->name);
	free(t);
}

//#define SCHED_DEBUG 1
static inline void call_post_monitor(struct sched *s, struct task *t)
{
	int ret;

#ifndef SCHED_DEBUG
	ret = t->info.post_monitor(s, t->info.context);
#else
	struct timeval t1, t2, diff;
	unsigned long pst;

	clock_get_realtime(&t1);
	ret = t->info.post_monitor(s, t->info.context);
	clock_get_realtime(&t2);
	tv_diff(&t1, &t2, &diff);
	pst = tv2ms(&diff);
	if (pst > 50)
		PARA_WARNING_LOG("%s: post_monitor time: %lums\n",
			t->name, pst);
#endif
	t->status = ret < 0? ret : TS_RUNNING;
}

static unsigned sched_post_monitor(struct sched *s)
{
	struct task *t, *tmp;
	unsigned num_running_tasks = 0;

	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
		if (t->status == TS_DEAD) /* task has been reaped */
			unlink_and_free_task(t);
		else if (t->status == TS_RUNNING) {
			call_post_monitor(s, t); /* sets t->status */
			t->notification = 0;
			if (t->status == TS_RUNNING)
				num_running_tasks++;
		}
	}
	return num_running_tasks;
}

/**
 * The core function of all paraslash programs.
 *
 * \param s Pointer to the scheduler struct.
 *
 * This function updates the global now pointer, calls all registered
 * pre_monitor hooks which may set the timeout and add any file descriptors to
 * the pollfd array. Next, it calls the poll function and makes the result
 * available to the registered tasks by calling their post_monitor hook.
 *
 * \return Zero if no more tasks are left in the task list, negative if the
 * poll function returned an error.
 *
 * \sa \ref now.
 */
int schedule(struct sched *s)
{
	int ret;
	unsigned num_running_tasks;

again:
	s->num_pfds = 0;
	if (s->pidx)
		memset(s->pidx, 0xff, s->pidx_array_len * sizeof(unsigned));
	s->timeout = s->default_timeout;
	clock_get_realtime(&now_struct);
	sched_pre_monitor(s);
	ret = s->poll_function(s->pfd, s->num_pfds, s->timeout);
	if (ret < 0)
		return ret;
	clock_get_realtime(&now_struct);
	num_running_tasks = sched_post_monitor(s);
	if (num_running_tasks == 0)
		return 0;
	goto again;
}

/**
 * Obtain the error status of a task and deallocate its resources.
 *
 * \param tptr Identifies the task to reap.
 *
 * When a task's post-monitor method returns negative, the task enters zombie
 * state. The {pre,post}-monitor methods will no longer be called, but some
 * amount of task-specific memory remains in use until this function is called.
 *
 * \return If tptr or *tptr is NULL, or if the given task is still running,
 * the function does nothing and returns zero. Otherwise it resets *tptr
 * and returns the (negative) error code of the terminated task.
 *
 * \sa \ref sched_shutdown(), wait(2).
 */
int task_reap(struct task **tptr)
{
	struct task *t;
	int ret;

	if (!tptr)
		return 0;
	t = *tptr;
	if (!t)
		return 0;
	if (t->status >= 0)
		return 0;
	ret = t->status;
	PARA_INFO_LOG("reaping %s: %s\n", t->name, para_strerror(-ret));
	/*
	 * With list_for_each_entry_safe() it is only safe to remove the
	 * _current_ list item. Since we are being called from the loop in
	 * schedule() via some task's ->post_monitor() function, freeing the
	 * given task here would result in use-after-free bugs in schedule().
	 * So we only set the task status to TS_DEAD which tells schedule() to
	 * free the task in the next iteration of its loop.
	 */
	t->status = TS_DEAD;

	*tptr = NULL;
	return ret;
}

/**
 * Deallocate all resources of all tasks of a scheduler instance.
 *
 * \param s The scheduler instance.
 *
 * This should only be called after \ref schedule() has returned.
 */
void sched_shutdown(struct sched *s)
{
	struct task *t, *tmp;

	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
		if (t->status == TS_RUNNING)
			/* The task list should contain only terminated tasks. */
			PARA_WARNING_LOG("shutting down running task %s\n",
				t->name);
		unlink_and_free_task(t);
	}
	free(s->pfd);
	free(s->pidx);
	free(s);
}

/**
 * Add a task to the scheduler task list.
 *
 * \param info Task information supplied by the caller.
 * \param s The scheduler instance.
 *
 * \return A pointer to a newly allocated task. The layout of the task
 * structure is only known to the scheduler core. The structure can be freed
 * after \ref schedule() returned ba calling \ref sched_shutdown().
 */
struct task *task_register(struct task_info *info, struct sched *s)
{
	struct task *t = alloc(sizeof(*t));

	assert(info->post_monitor);
	t->info = *info;
	t->name = para_strdup(info->name);
	t->notification = 0;
	t->status = TS_RUNNING;
	list_add_tail(&t->node, &s->task_list);
	return t;
}

/**
 * Get the list of all registered tasks.
 *
 * \param s The scheduler instance to get the task list from.
 *
 * \return The task list.
 *
 * Each entry of the list contains an identifier which is simply a hex number.
 * The result is dynamically allocated and must be freed by the caller.
 */
char *get_task_list(struct sched *s)
{
	struct task *t, *tmp;
	char *msg = NULL;

	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
		char *tmp_msg;
		tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
			t->status == TS_DEAD? "dead" :
				(t->status == TS_RUNNING? "running" : "zombie"),
			t->name);
		free(msg);
		msg = tmp_msg;
	}
	return msg;
}

/**
 * Set the notification value of a task.
 *
 * \param t The task to notify.
 * \param err A positive error code.
 *
 * Tasks which honor notifications are supposed to call \ref
 * task_get_notification() in their post_monitor function and act on the
 * returned notification value.
 *
 * If the scheduler detects during its pre_monitor loop that at least one task
 * has been notified, the loop terminates, and the post_monitor methods of all
 * taks are immediately called again.
 *
 * The notification for a task is reset after the call to its post_monitor
 * method.
 *
 * \sa \ref task_get_notification().
 */
void task_notify(struct task *t, int err)
{
	assert(err > 0);
	if (t->notification == -err) /* ignore subsequent notifications */
		return;
	PARA_INFO_LOG("notifying task %s: %s\n", t->name, para_strerror(err));
	t->notification = -err;
}

/**
 * Return the notification value of a task.
 *
 * \param t The task to get the notification value from.
 *
 * \return The notification value. If this is negative, the task has been
 * notified by another task. Tasks are supposed to check for notifications by
 * calling this function from their post_monitor method.
 *
 * \sa \ref task_notify().
 */
int task_get_notification(const struct task *t)
{
	return t->notification;
}

/**
 * Return the status value of a task.
 *
 * \param t The task to get the status value from.
 *
 * \return Zero if task does not exist, one if task is running, negative error
 * code if task has terminated.
 */
int task_status(const struct task *t)
{
	if (!t)
		return 0;
	if (t->status == TS_DEAD) /* pretend dead tasks don't exist */
		return 0;
	if (t->status == TS_RUNNING)
		return 1;
	return t->status;
}

/**
 * Set the notification value of all tasks of a scheduler instance.
 *
 * \param s The scheduler instance whose tasks should be notified.
 * \param err A positive error code.
 *
 * This simply iterates over all existing tasks of \a s and sets each
 * task's notification value to \p -err.
 */
void task_notify_all(struct sched *s, int err)
{
	struct task *t;

	list_for_each_entry(t, &s->task_list, node)
		task_notify(t, err);
}

/**
 * Set the I/O timeout to the minimal possible value.
 *
 * \param s Pointer to the scheduler struct.
 *
 * This causes the next poll() call to return immediately.
 */
void sched_min_delay(struct sched *s)
{
	s->timeout = 0;
}

/**
 * Impose an upper bound for the I/O timeout.
 *
 * \param to Maximal allowed timeout.
 * \param s Pointer to the scheduler struct.
 *
 * If the current I/O timeout is already smaller than to, this function does
 * nothing. Otherwise the timeout is set to the given value.
 *
 * \sa \ref sched_request_timeout_ms().
 */
void sched_request_timeout(struct timeval *to, struct sched *s)
{
	long unsigned ms = tv2ms(to);
	if (s->timeout > ms)
		s->timeout = ms;
}

/**
 * Bound the I/O timeout to at most the given amount of milliseconds.
 *
 * \param ms The maximal allowed timeout in milliseconds.
 * \param s Pointer to the scheduler struct.
 *
 * Like \ref sched_request_timeout() this imposes an upper bound on the I/O
 * timeout.
 */
void sched_request_timeout_ms(long unsigned ms, struct sched *s)
{
	struct timeval tv;
	ms2tv(ms, &tv);
	sched_request_timeout(&tv, s);
}

/**
 * Bound the I/O timeout by an absolute time in the future.
 *
 * \param barrier Defines the upper bound for the timeout.
 * \param s Pointer to the scheduler struct.
 *
 * \return If the barrier is in the past, this function does nothing and
 * returns zero. Otherwise it returns one.
 *
 * \sa \ref sched_request_barrier_or_min_delay().
 */
int sched_request_barrier(struct timeval *barrier, struct sched *s)
{
	struct timeval diff;

	if (tv_diff(now, barrier, &diff) > 0)
		return 0;
	sched_request_timeout(&diff, s);
	return 1;
}

/**
 * Bound the I/O timeout or request a minimal delay.
 *
 * \param barrier Absolute time as in \ref sched_request_barrier().
 * \param s Pointer to the scheduler struct.
 *
 * \return If the barrier is in the past, this function requests a minimal
 * timeout and returns zero. Otherwise it returns one.
 *
 * \sa \ref sched_min_delay(), \ref sched_request_barrier().
 */
int sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
{
	struct timeval diff;

	if (tv_diff(now, barrier, &diff) > 0) {
		sched_min_delay(s);
		return 0;
	}
	sched_request_timeout(&diff, s);
	return 1;
}

static void add_pollfd(int fd, struct sched *s, short events)
{
	assert(fd >= 0);
#if 0
	{
		int flags = fcntl(fd, F_GETFL);
		if (!(flags & O_NONBLOCK)) {
			PARA_EMERG_LOG("fd %d is a blocking file descriptor\n", fd);
			exit(EXIT_FAILURE);
		}
	}
#endif
	if (s->pidx_array_len > fd) { /* is fd already registered? */
		if (s->pidx[fd] < s->pfd_array_len) { /* yes, it is */
			assert(s->pfd[s->pidx[fd]].fd == fd);
			s->pfd[s->pidx[fd]].events |= events;
			return;
		}
	} else { /* need to extend the index array */
		unsigned old_len = s->pidx_array_len;
		while (s->pidx_array_len <= fd)
			s->pidx_array_len = s->pidx_array_len * 2 + 1;
		PARA_INFO_LOG("pidx array len: %u\n", s->pidx_array_len);
		s->pidx = para_realloc(s->pidx,
			s->pidx_array_len * sizeof(unsigned));
		memset(s->pidx + old_len, 0xff,
			(s->pidx_array_len - old_len) * sizeof(unsigned));
	}
	/*
	 * The given fd is not part of the pfd array yet. Initialize pidx[fd]
	 * to point at the next unused slot of this array and initialize the
	 * slot.
	 */
	s->pidx[fd] = s->num_pfds;
	if (s->pfd_array_len <= s->num_pfds) {
		unsigned old_len = s->pfd_array_len;
		s->pfd_array_len = old_len * 2 + 1;
		PARA_INFO_LOG("pfd array len: %u\n", s->pfd_array_len);
		s->pfd = para_realloc(s->pfd,
			s->pfd_array_len * sizeof(struct pollfd));
		memset(s->pfd + old_len, 0,
			(s->pfd_array_len - old_len) * sizeof(struct pollfd));
	}
	s->pfd[s->num_pfds].fd = fd;
	s->pfd[s->num_pfds].events = events;
	s->pfd[s->num_pfds].revents = 0;
	s->num_pfds++;
}

/**
 * Instruct the scheduler to monitor an fd for readiness for reading.
 *
 * \param fd The file descriptor.
 * \param s The scheduler.
 *
 * \sa \ref sched_monitor_writefd().
 */
void sched_monitor_readfd(int fd, struct sched *s)
{
	add_pollfd(fd, s, POLLIN);
}

/**
 * Instruct the scheduler to monitor an fd for readiness for writing.
 *
 * \param fd The file descriptor.
 * \param s The scheduler.
 *
 * \sa \ref sched_monitor_readfd().
 */
void sched_monitor_writefd(int fd, struct sched *s)
{
	add_pollfd(fd, s, POLLOUT);
}

static int get_revents(int fd, const struct sched *s)
{
	if (fd < 0)
		return 0;
	if (fd >= s->pidx_array_len)
		return 0;
	if (s->pidx[fd] >= s->num_pfds)
		return 0;
	if (s->pfd[s->pidx[fd]].fd != fd)
		return 0;
	assert((s->pfd[s->pidx[fd]].revents & POLLNVAL) == 0);
	return s->pfd[s->pidx[fd]].revents;
}

/**
 * Check whether there is data to read on the given fd.
 *
 * To be called from the ->post_monitor() method of a task.
 *
 * \param fd Should have been monitored with \ref sched_monitor_readfd().
 * \param s The scheduler instance.
 *
 * \return True if the file descriptor is ready for reading, false otherwise.
 * If fd is negative, or has not been monitored in the current iteration of the
 * scheduler's main loop, the function also returns false.
 *
 * \sa \ref sched_write_ok().
 */
bool sched_read_ok(int fd, const struct sched *s)
{
	return get_revents(fd, s) & (POLLIN | POLLERR | POLLHUP);
}

/**
 * Check whether writing is possible (i.e., does not block).
 *
 * \param fd Should have been monitored with \ref sched_monitor_writefd().
 * \param s The scheduler instance.
 *
 * \return True if the file descriptor is ready for writing, false otherwise.
 * The comment in \ref sched_read_ok() about invalid file descriptors applies
 * to this function as well.
 */
bool sched_write_ok(int fd, const struct sched *s)
{
	return get_revents(fd, s) & (POLLOUT | POLLERR | POLLHUP);
}