Web/reactor/epoll.c

#include "epoll.h"

#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <errno.h>

// fd_data holds per-file-descriptor data for O(1) dispatch
typedef struct fd_data {
	handler_fn handler;    // callback for this FD
	void *user_data;       // arbitrary user data passed to handler
} fd_data_t;

// runtime manages the event loop
struct runtime {
	int epfd;                      // epoll file descriptor
	int max_events;                // max events per epoll_wait
	struct epoll_event *events;    // buffer for epoll_wait results
	fd_data_t *handlers;           // array indexed by FD
	int max_fd;                    // current size of handlers array
	bool running;                  // controls event loop
};

// runtime_new creates a new event loop runtime.
// max_events controls batching - higher values mean fewer syscalls but more latency.
// Typical values are 128-1024.
runtime_t* runtime_new(int max_events) {
	// Allocate runtime structure with all fields zeroed
	runtime_t *rt = calloc(1, sizeof(runtime_t));
	if (!rt) return NULL;

	// Create epoll instance with CLOEXEC for exec() safety
	rt->epfd = epoll_create1(EPOLL_CLOEXEC);
	if (rt->epfd < 0) {
		free(rt);
		return NULL;
	}

	// Allocate event buffer for epoll_wait results
	rt->max_events = max_events;
	rt->events = calloc(max_events, sizeof(struct epoll_event));
	if (!rt->events) {
		close(rt->epfd);
		free(rt);
		return NULL;
	}

	// Initialize handler array with reasonable initial size
	// This will grow dynamically as needed
	rt->max_fd = 1024;
	rt->handlers = calloc(rt->max_fd, sizeof(fd_data_t));
	if (!rt->handlers) {
		free(rt->events);
		close(rt->epfd);
		free(rt);
		return NULL;
	}

	rt->running = false;
	return rt;
}

// runtime_free destroys a runtime and frees all resources.
// Does NOT close user file descriptors - caller is responsible for those.
void runtime_free(runtime_t *rt) {
	if (!rt) return;
	close(rt->epfd);
	free(rt->events);
	free(rt->handlers);
	free(rt);
}

// ensure_capacity grows the handlers array if needed.
// The handlers array is indexed by FD for O(1) lookup.
// Growth is done in chunks of 1024 to minimize reallocations.
static int ensure_capacity(runtime_t *rt, int fd) {
	if (fd >= rt->max_fd) {
		// Grow by 1024 FDs at a time to reduce realloc frequency
		int new_max = fd + 1024;
		fd_data_t *new_handlers = realloc(rt->handlers, new_max * sizeof(fd_data_t));
		if (!new_handlers) return -1;

		// Zero out the new portion for NULL handler checks
		memset(new_handlers + rt->max_fd, 0, (new_max - rt->max_fd) * sizeof(fd_data_t));
		rt->handlers = new_handlers;
		rt->max_fd = new_max;
	}
	return 0;
}

// runtime_add adds a file descriptor to the event loop.
// Common event flags:
// - EPOLLIN: Ready for reading
// - EPOLLOUT: Ready for writing
// - EPOLLET: Edge-triggered mode (recommended for performance)
// - EPOLLONESHOT: Disable after one event (useful for thread pools)
int runtime_add(runtime_t *rt, int fd, uint32_t events, handler_fn handler, void *data) {
	// Make sure our handler array is big enough
	if (ensure_capacity(rt, fd) < 0) return -1;

	// Configure epoll_event - store FD for easy retrieval
	struct epoll_event ev = {
		.events = events,
		.data.fd = fd
	};

	// Register with epoll
	if (epoll_ctl(rt->epfd, EPOLL_CTL_ADD, fd, &ev) < 0) {
		return -1;
	}

	// Store handler and user data for O(1) dispatch
	rt->handlers[fd].handler = handler;
	rt->handlers[fd].user_data = data;
	return 0;
}

// runtime_mod modifies events and/or handler for an existing FD.
// Useful for switching between read/write modes or changing handlers.
int runtime_mod(runtime_t *rt, int fd, uint32_t events, handler_fn handler, void *data) {
	if (ensure_capacity(rt, fd) < 0) return -1;

	struct epoll_event ev = {
		.events = events,
		.data.fd = fd
	};

	// Update epoll registration
	if (epoll_ctl(rt->epfd, EPOLL_CTL_MOD, fd, &ev) < 0) {
		return -1;
	}

	// Update handler storage
	rt->handlers[fd].handler = handler;
	rt->handlers[fd].user_data = data;
	return 0;
}

// runtime_del removes a file descriptor from the event loop.
// Does NOT close the FD - caller is responsible for that.
int runtime_del(runtime_t *rt, int fd) {
	// Remove from epoll (NULL event allowed on Linux >= 2.6.9)
	if (epoll_ctl(rt->epfd, EPOLL_CTL_DEL, fd, NULL) < 0) {
		return -1;
	}

	// Clear handler storage if FD is in range
	if (fd < rt->max_fd) {
		rt->handlers[fd].handler = NULL;
		rt->handlers[fd].user_data = NULL;
	}
	return 0;
}

// runtime_run starts the event loop.
// Blocks and dispatches handlers as events occur until runtime_stop() is called.
// Handlers receive: runtime, fd, events mask, and their user data.
int runtime_run(runtime_t *rt) {
	rt->running = true;

	while (rt->running) {
		// Wait for events. -1 timeout means block indefinitely
		int n = epoll_wait(rt->epfd, rt->events, rt->max_events, -1);

		if (n < 0) {
			// EINTR happens on signals - just retry
			if (errno == EINTR) continue;
			return -1;
		}

		// Process all ready file descriptors
		for (int i = 0; i < n; i++) {
			int fd = rt->events[i].data.fd;
			uint32_t events = rt->events[i].events;

			// Bounds check and NULL check before dispatch
			// FDs can be removed by handlers, so we must check
			if (fd < rt->max_fd && rt->handlers[fd].handler) {
				rt->handlers[fd].handler(rt, fd, events, rt->handlers[fd].user_data);
			}
		}
	}

	return 0;
}

// runtime_stop signals the event loop to stop.
// Can be called from signal handlers or event handlers for clean shutdown.
void runtime_stop(runtime_t *rt) {
	rt->running = false;
}

// set_nonblocking makes a file descriptor non-blocking.
// Essential for epoll servers - blocking I/O would freeze the entire event loop.
int set_nonblocking(int fd) {
	// Get current flags
	int flags = fcntl(fd, F_GETFL, 0);
	if (flags < 0) return -1;

	// Add O_NONBLOCK flag
	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
}