fix: transparently support blocking work without worker threads

example/blocking.c

@@ -27,12 +27,12 @@ int main() {
                      ctx_reset,
                      ctx_free);
 
-  lnm_http_step_init(&step, slow_step);
+  lnm_http_step_append(&step, slow_step, true);
   lnm_http_route_init_literal(&route, lnm_http_method_get, "/", step);
   lnm_http_loop_route_add(hl, route);
 
   lnm_log_init_global();
   lnm_log_register_stdout(lnm_log_level_debug);
 
-  lnm_http_loop_run(hl, 8080, 1);
+  printf("res = %i\n", lnm_http_loop_run(hl, 8080, 1, 2));
 }

include/lnm/http/loop.h

@@ -32,23 +32,16 @@ lnm_err lnm_http_loop_init(lnm_http_loop **out, void *c_gctx,
                            lnm_http_ctx_reset_fn ctx_reset,
                            lnm_http_ctx_free_fn ctx_free);
 
-/**
- * Initialize a new step.
- *
- * @param out where to store pointer to new `lnm_http_step`
- * @param fn step function
- */
-lnm_err lnm_http_step_init(lnm_http_step **out, lnm_http_step_fn fn);
-
 /**
  * Append the given step fn to the step.
  *
- * @param out where to store pointer to new `lnm_http_step`
- * @param step step to append new step to
+ * @param out both the previous step to append the new step to, and the output
+ * variable to which the new step is appended
  * @param fn step function
+ * @param blocking whether the step is blocking or not
  */
-lnm_err lnm_http_step_append(lnm_http_step **out, lnm_http_step *step,
-                             lnm_http_step_fn fn);
+lnm_err lnm_http_step_append(lnm_http_step **out, lnm_http_step_fn fn,
+                             bool blocking);
 
 /**
  * Initialize a new route of type literal.
@@ -81,7 +74,8 @@ lnm_err lnm_http_route_init_regex(lnm_http_route **out, lnm_http_method method,
  */
 lnm_err lnm_http_loop_route_add(lnm_http_loop *hl, lnm_http_route *route);
 
-lnm_err lnm_http_loop_run(lnm_http_loop *hl, uint16_t port, int thread_count);
+lnm_err lnm_http_loop_run(lnm_http_loop *hl, uint16_t port,
+                          size_t epoll_threads, size_t worker_threads);
 
 void lnm_http_loop_set_api_key(lnm_http_loop *hl, const char *api_key);
 

include/lnm/loop.h

@@ -1,6 +1,7 @@
 #ifndef LNM_LOOP
 #define LNM_LOOP
 
+#include <pthread.h>
 #include <stdatomic.h>
 #include <stdint.h>
 #include <stdlib.h>
@@ -8,13 +9,19 @@
 #include "lnm/common.h"
 
 #define LNM_LOOP_BUF_SIZE 2048
+#define LNM_QUEUE_MULTIPLIER 8
 
-typedef enum {
-  lnm_loop_state_req = 0,
-  lnm_loop_state_res,
+typedef enum lnm_loop_state {
+  lnm_loop_state_req_io = 0,
+  lnm_loop_state_res_io,
   lnm_loop_state_end,
+  lnm_loop_state_req_work,
+  lnm_loop_state_res_work,
 } lnm_loop_state;
 
+/**
+ * State for a currently active connection
+ */
 typedef struct lnm_loop_conn {
   int fd;
   lnm_loop_state state;
@@ -30,6 +37,38 @@ typedef struct lnm_loop_conn {
   } w;
 } lnm_loop_conn;
 
+/**
+ * Concurrent fixed-size queue used to distribute work among worker threads
+ */
+typedef struct lnm_loop_queue {
+  struct {
+    lnm_loop_conn **arr;
+    size_t len;
+  } buf;
+  size_t head;
+  size_t tail;
+  bool empty;
+  pthread_mutex_t mutex;
+  pthread_cond_t cond;
+} lnm_loop_queue;
+
+/**
+ * Initialize a new queue with the specified fixed capacity
+ */
+lnm_err lnm_loop_queue_init(lnm_loop_queue **out, size_t cap);
+
+/**
+ * Queue the given connection. If the queue is currently full, this action
+ * blocks until there is space available.
+ */
+void lnm_loop_queue_push(lnm_loop_queue *q, lnm_loop_conn *conn);
+
+/**
+ * Pop a connection from the queue. This action blocks until a connection is
+ * available.
+ */
+lnm_loop_conn *lnm_loop_queue_pop(lnm_loop_queue *q);
+
 typedef struct lnm_loop {
   int listen_fd;
   int epoll_fd;
@@ -39,6 +78,13 @@ typedef struct lnm_loop {
   void (*ctx_free)(void *ctx);
   void (*data_read)(lnm_loop_conn *conn);
   void (*data_write)(lnm_loop_conn *conn);
+  lnm_loop_queue *wq;
+  struct {
+    // Mutex shared between all threads; used for counting thread IDs
+    pthread_mutex_t mutex;
+    size_t worker_count;
+    size_t epoll_count;
+  } threads;
 } lnm_loop;
 
 lnm_err lnm_loop_init(lnm_loop **out, void *gctx,
@@ -49,6 +95,46 @@ lnm_err lnm_loop_init(lnm_loop **out, void *gctx,
 
 lnm_err lnm_loop_setup(lnm_loop *l, uint16_t port);
 
-lnm_err lnm_loop_run(lnm_loop *l, int thread_count);
+/**
+ * Run a single epoll thread of the event loop.
+ */
+lnm_err lnm_loop_run(lnm_loop *l);
+
+/**
+ * Run a multithreaded event loop with the configured number of threads.
+ */
+lnm_err lnm_loop_run_multi(lnm_loop *l, size_t epoll_threads,
+                           size_t worker_threads);
+
+/**
+ * Advance the processing of the given connection.
+ *
+ * Behavior of this function depends on both the connection state and whether
+ * worker threads are enabled.
+ *
+ * For IO states, this function will perform network I/O along with executing
+ * the loop's respective processing functions.
+ *
+ * For work states, the respective processing functions are executed without
+ * performing any network I/O. If no worker queue is present, this function
+ * performs all blocking work until an I/O or the end state is reached. If there
+ * is a worker queue present, only one block of work is done before exiting,
+ * allowing further blocks of work to be scheduled on other worker threads.
+ *
+ * If no worker queue is present, this function will only exit once an I/O or
+ * end state is reached.
+ */
+void lnm_loop_conn_advance(lnm_loop *l, lnm_loop_conn *conn);
+
+/**
+ * Reschedule the given connection, either on the event loop for network IO or
+ * on a worker thread for blocking work. Connections are terminated as needed.
+ */
+void lnm_loop_conn_schedule(lnm_loop *l, lnm_loop_conn *conn);
+
+/**
+ * Main loop executed on the worker threads.
+ */
+void lnm_loop_worker_run(void *arg);
 
 #endif

src/_include/lnm/http/loop_internal.h

@@ -8,6 +8,7 @@
 typedef struct lnm_http_step {
   lnm_http_step_fn fn;
   struct lnm_http_step *next;
+  bool blocking;
 } lnm_http_step;
 
 typedef enum lnm_http_route_type {

src/_include/lnm/loop_internal.h

@@ -9,6 +9,6 @@ void lnm_loop_conn_free(lnm_loop *l, lnm_loop_conn *conn);
 
 lnm_err lnm_loop_accept(lnm_loop *l);
 
-void lnm_loop_conn_io(lnm_loop *l, lnm_loop_conn *conn);
+void lnm_loop_conn_advance(lnm_loop *l, lnm_loop_conn *conn);
 
 #endif

src/http/lnm_http_loop.c

@@ -28,7 +28,8 @@ lnm_err lnm_http_loop_init(lnm_http_loop **out, void *c_gctx,
   return lnm_err_ok;
 }
 
-lnm_err lnm_http_step_init(lnm_http_step **out, lnm_http_step_fn fn) {
+lnm_err lnm_http_step_append(lnm_http_step **out, lnm_http_step_fn fn,
+                             bool blocking) {
   lnm_http_step *step = calloc(1, sizeof(lnm_http_step));
 
   if (step == NULL) {
@@ -36,19 +37,14 @@ lnm_err lnm_http_step_init(lnm_http_step **out, lnm_http_step_fn fn) {
   }
 
   step->fn = fn;
-  *out = step;
+  step->blocking = blocking;
 
-  return lnm_err_ok;
-}
-
-lnm_err lnm_http_step_append(lnm_http_step **out, lnm_http_step *step,
-                             lnm_http_step_fn fn) {
-  LNM_RES(lnm_http_step_init(out, fn));
-
-  if (step != NULL) {
-    step->next = *out;
+  if ((*out) != NULL) {
+    (*out)->next = step;
   }
 
+  *out = step;
+
   return lnm_err_ok;
 }
 
@@ -122,9 +118,10 @@ lnm_err lnm_http_loop_route_add(lnm_http_loop *hl, lnm_http_route *route) {
   return lnm_err_ok;
 }
 
-lnm_err lnm_http_loop_run(lnm_http_loop *hl, uint16_t port, int thread_count) {
+lnm_err lnm_http_loop_run(lnm_http_loop *hl, uint16_t port,
+                          size_t epoll_threads, size_t worker_threads) {
   LNM_RES(lnm_loop_setup(hl, port));
-  return lnm_loop_run(hl, thread_count);
+  return lnm_loop_run_multi(hl, epoll_threads, worker_threads);
 }
 
 void lnm_http_loop_set_api_key(lnm_http_loop *hl, const char *api_key) {

src/http/lnm_http_loop_process.c

@@ -133,11 +133,19 @@ void lnm_http_loop_process_steps(lnm_http_conn *conn) {
   while ((ctx->cur_step != NULL) && (step != ctx->cur_step)) {
     step = ctx->cur_step;
 
+    if (step->blocking && (conn->state != lnm_loop_state_req_work)) {
+      conn->state = lnm_loop_state_req_work;
+
+      break;
+    }
+
     switch (step->fn(conn)) {
     case lnm_http_step_err_done:
       ctx->cur_step = ctx->cur_step->next;
       break;
     case lnm_http_step_err_io_needed:
+      // Ensure steps that require more I/O are executed on the event loop
+      conn->state = lnm_loop_state_req_io;
       break;
     case lnm_http_step_err_close:
       conn->state = lnm_loop_state_end;
@@ -149,6 +157,7 @@ void lnm_http_loop_process_steps(lnm_http_conn *conn) {
   }
 
   if (ctx->cur_step == NULL) {
+    conn->state = lnm_loop_state_res_io;
     ctx->state = lnm_http_loop_state_add_headers;
   }
 }
@@ -325,23 +334,23 @@ void (*process_fns[])(lnm_http_conn *conn) = {
 
 lnm_loop_state state_map[] = {
     // parse_req
-    lnm_loop_state_req,
+    lnm_loop_state_req_io,
     // route
-    lnm_loop_state_req,
+    lnm_loop_state_req_io,
     // parse_headers
-    lnm_loop_state_req,
+    lnm_loop_state_req_io,
     // steps
-    lnm_loop_state_req,
+    lnm_loop_state_req_io,
     // add_headers
-    lnm_loop_state_req,
+    lnm_loop_state_req_io,
     // write_status_line
-    lnm_loop_state_res,
+    lnm_loop_state_res_io,
     // write_headers
-    lnm_loop_state_res,
+    lnm_loop_state_res_io,
     // write_body
-    lnm_loop_state_res,
+    lnm_loop_state_res_io,
     // finish
-    lnm_loop_state_res,
+    lnm_loop_state_res_io,
 };
 
 void lnm_http_loop_process(lnm_http_conn *conn) {
@@ -369,7 +378,7 @@ void lnm_http_loop_process(lnm_http_conn *conn) {
 
   // We move the request to a dedicated buffer if the read buffer needs to be
   // reused
-  if ((conn->state == lnm_loop_state_req) && (conn->state == loop_state) &&
+  if ((conn->state == lnm_loop_state_req_io) && (conn->state == loop_state) &&
       (!ctx->req.buf.owned) && (ctx->req.buf.len > 0)) {
     char *buf = malloc(ctx->req.buf.len);
 

src/loop/lnm_loop.c

@@ -30,6 +30,8 @@ lnm_err lnm_loop_init(lnm_loop **out, void *gctx,
   l->data_read = data_read;
   l->data_write = data_write;
 
+  pthread_mutex_init(&l->threads.mutex, NULL);
+
   *out = l;
 
   return lnm_err_ok;
@@ -53,7 +55,7 @@ lnm_err lnm_loop_accept(lnm_loop *l) {
   LNM_RES2(lnm_loop_conn_init(&conn, l), close(conn_fd));
 
   conn->fd = conn_fd;
-  conn->state = lnm_loop_state_req;
+  conn->state = lnm_loop_state_req_io;
 
   struct epoll_event event = {.data.ptr = conn,
                               .events = EPOLLIN | EPOLLET | EPOLLONESHOT};
@@ -62,6 +64,10 @@ lnm_err lnm_loop_accept(lnm_loop *l) {
 
   l->open++;
 
+  // Make sure to re-arm the listening socket after accepting
+  event.data.ptr = NULL;
+  epoll_ctl(l->epoll_fd, EPOLL_CTL_MOD, l->listen_fd, &event);
+
   lnm_ldebug(section, "connection opened with fd %i", conn_fd);
 
   return lnm_err_ok;
@@ -123,19 +129,51 @@ lnm_err lnm_loop_setup(lnm_loop *l, uint16_t port) {
   return lnm_err_ok;
 }
 
-typedef struct lnm_loop_thread_args {
-  lnm_loop *l;
-  int id;
-  int thread_count;
-} lnm_loop_thread_args;
+void lnm_loop_conn_schedule(lnm_loop *l, lnm_loop_conn *conn) {
+  switch (conn->state) {
+  // IO states get rescheduled in the epoll loop
+  case lnm_loop_state_req_io:
+  case lnm_loop_state_res_io: {
+    struct epoll_event event = {
+        .data.ptr = conn,
+        .events = (conn->state == lnm_loop_state_req_io ? EPOLLIN : EPOLLOUT) |
+                  EPOLLET | EPOLLONESHOT};
 
-lnm_err lnm_loop_run_thread(lnm_loop_thread_args *args) {
-  lnm_loop *l = args->l;
-  int thread_id = args->id;
-  int thread_count = args->thread_count;
+    epoll_ctl(l->epoll_fd, EPOLL_CTL_MOD, conn->fd, &event);
+  } break;
+  case lnm_loop_state_req_work:
+  case lnm_loop_state_res_work:
+    lnm_loop_queue_push(l->wq, conn);
+    break;
+  case lnm_loop_state_end: {
+    int conn_fd = conn->fd;
+
+    lnm_loop_conn_free(l, conn);
+    close(conn_fd);
+    l->open--;
+
+    epoll_ctl(l->epoll_fd, EPOLL_CTL_DEL, conn_fd, NULL);
+
+    lnm_ldebug(section, "connection closed with fd %i", conn_fd);
+  } break;
+  }
+}
+
+lnm_err lnm_loop_run(lnm_loop *l) {
+  if (l->epoll_fd == 0) {
+    return lnm_err_not_setup;
+  }
+
+  // Get thread ID by incrementing counter
+  pthread_mutex_lock(&l->threads.mutex);
+
+  int thread_id = l->threads.epoll_count;
+  l->threads.epoll_count++;
+
+  pthread_mutex_unlock(&l->threads.mutex);
 
   struct epoll_event *events = calloc(1, sizeof(struct epoll_event));
-  int events_cap = 1;
+  size_t events_cap = 1;
 
   if (events == NULL) {
     return lnm_err_failed_alloc;
@@ -143,9 +181,6 @@ lnm_err lnm_loop_run_thread(lnm_loop_thread_args *args) {
 
   lnm_linfo(section, "thread %i started", thread_id);
 
-  struct epoll_event listen_event = {
-      .data.ptr = NULL, .events = EPOLLIN | EPOLLET | EPOLLONESHOT};
-
   while (1) {
     int polled = epoll_wait(l->epoll_fd, events, events_cap, -1);
     lnm_ldebug(section, "polled (thread %i): %i", thread_id, polled);
@@ -157,36 +192,19 @@ lnm_err lnm_loop_run_thread(lnm_loop_thread_args *args) {
     for (int i = 0; i < polled; i++) {
       if (events[i].data.ptr == NULL) {
         lnm_loop_accept(l);
-
-        epoll_ctl(l->epoll_fd, EPOLL_CTL_MOD, l->listen_fd, &listen_event);
       } else {
         lnm_loop_conn *conn = events[i].data.ptr;
-        lnm_loop_conn_io(l, conn);
 
-        if (conn->state == lnm_loop_state_end) {
-          int conn_fd = conn->fd;
-
-          lnm_loop_conn_free(l, conn);
-          close(conn_fd);
-          l->open--;
-
-          epoll_ctl(l->epoll_fd, EPOLL_CTL_DEL, conn_fd, NULL);
-
-          lnm_ldebug(section, "connection closed with fd %i", conn_fd);
-        } else {
-          struct epoll_event event = {
-              .data.ptr = conn,
-              .events =
-                  (conn->state == lnm_loop_state_req ? EPOLLIN : EPOLLOUT) |
-                  EPOLLET | EPOLLONESHOT};
-          epoll_ctl(l->epoll_fd, EPOLL_CTL_MOD, conn->fd, &event);
-        }
+        // At this point, state is always an IO state
+        lnm_loop_conn_advance(l, conn);
+        lnm_loop_conn_schedule(l, conn);
       }
     }
 
-    int open = l->open;
-    int cap_per_thread =
-        open + 1 > thread_count ? (open + 1) / thread_count : 1;
+    size_t open = l->open;
+    size_t cap_per_thread = open + 1 > l->threads.epoll_count
+                                ? (open + 1) / l->threads.epoll_count
+                                : 1;
 
     if (cap_per_thread > events_cap) {
       struct epoll_event *new_events =
@@ -205,28 +223,21 @@ lnm_err lnm_loop_run_thread(lnm_loop_thread_args *args) {
   return lnm_err_ok;
 }
 
-lnm_err lnm_loop_run(lnm_loop *l, int thread_count) {
-  if (l->epoll_fd == 0) {
-    return lnm_err_not_setup;
+lnm_err lnm_loop_run_multi(lnm_loop *l, size_t epoll_threads,
+                           size_t worker_threads) {
+  if (worker_threads > 0) {
+    LNM_RES(lnm_loop_queue_init(&l->wq, LNM_QUEUE_MULTIPLIER * worker_threads));
   }
 
-  lnm_loop_thread_args args[thread_count];
+  pthread_t t;
 
-  for (int i = 1; i < thread_count; i++) {
-    args[i].l = l;
-    args[i].id = i;
-    args[i].thread_count = thread_count;
-
-    pthread_t thread;
-    pthread_create(&thread, NULL, (void *(*)(void *))lnm_loop_run_thread,
-                   &args[i]);
+  for (size_t i = 1; i < epoll_threads; i++) {
+    pthread_create(&t, NULL, (void *(*)(void *))lnm_loop_run, l);
   }
 
-  args[0].l = l;
-  args[0].id = 0;
-  args[0].thread_count = thread_count;
+  for (size_t i = 0; i < worker_threads; i++) {
+    pthread_create(&t, NULL, (void *(*)(void *))lnm_loop_worker_run, l);
+  }
 
-  lnm_loop_run_thread(&args[0]);
-
-  return lnm_err_ok;
+  return lnm_loop_run(l);
 }

src/loop/lnm_loop_io.c

@@ -1,5 +1,6 @@
 #include <errno.h>
 #include <string.h>
+#include <sys/socket.h>
 #include <unistd.h>
 
 #include "lnm/loop.h"
@@ -33,7 +34,7 @@ void lnm_loop_conn_io_req(lnm_loop *l, lnm_loop_conn *conn) {
 
     conn->r.size += res;
     l->data_read(conn);
-  } while (conn->state == lnm_loop_state_req);
+  } while (conn->state == lnm_loop_state_req_io);
 }
 
 void lnm_loop_conn_io_res(lnm_loop *l, lnm_loop_conn *conn) {
@@ -43,7 +44,9 @@ void lnm_loop_conn_io_res(lnm_loop *l, lnm_loop_conn *conn) {
     ssize_t res;
 
     do {
-      res = write(conn->fd, conn->w.buf, conn->w.size);
+      // Send with MSG_NOSIGNAL prevents closed pipes from exiting the program
+      // with SIGPIPE
+      res = send(conn->fd, conn->w.buf, conn->w.size, MSG_NOSIGNAL);
     } while (res < 0 && errno == EINTR);
 
     // Write can't be performed without blocking; we come back later
@@ -61,17 +64,33 @@ void lnm_loop_conn_io_res(lnm_loop *l, lnm_loop_conn *conn) {
     // writer function more space to work with
     memmove(conn->w.buf, &conn->w.buf[res], conn->w.size - res);
     conn->w.size -= res;
-  } while (conn->state == lnm_loop_state_res);
+  } while (conn->state == lnm_loop_state_res_io);
 }
 
-void lnm_loop_conn_io(lnm_loop *l, lnm_loop_conn *conn) {
-  switch (conn->state) {
-  case lnm_loop_state_req:
-    lnm_loop_conn_io_req(l, conn);
-    break;
-  case lnm_loop_state_res:
-    lnm_loop_conn_io_res(l, conn);
-    break;
-  default:;
+void lnm_loop_conn_advance(lnm_loop *l, lnm_loop_conn *conn) {
+  do {
+
+    switch (conn->state) {
+    case lnm_loop_state_req_io:
+      lnm_loop_conn_io_req(l, conn);
+      break;
+    case lnm_loop_state_res_io:
+      lnm_loop_conn_io_res(l, conn);
+      break;
+    case lnm_loop_state_req_work:
+      do {
+        l->data_read(conn);
+      } while (conn->state == lnm_loop_state_req_work);
+      break;
+    case lnm_loop_state_res_work:
+      do {
+        l->data_write(conn);
+      } while (conn->state == lnm_loop_state_res_work);
+      break;
+    default:;
+    }
   }
+  // Execute all blocking work if we're running in single-threaded mode
+  while (l->wq == NULL && (conn->state == lnm_loop_state_req_work ||
+                           conn->state == lnm_loop_state_res_work));
 }

src/loop/lnm_loop_worker.c

@@ -0,0 +1,92 @@
+#include <sys/epoll.h>
+
+#include "lnm/log.h"
+#include "lnm/loop.h"
+
+lnm_err lnm_loop_queue_init(lnm_loop_queue **out, size_t cap) {
+  lnm_loop_conn **arr = calloc(cap, sizeof(lnm_loop_conn *));
+
+  if (arr == NULL) {
+    return lnm_err_failed_alloc;
+  }
+
+  lnm_loop_queue *q = calloc(1, sizeof(lnm_loop_queue));
+
+  if (q == NULL) {
+    free(arr);
+
+    return lnm_err_failed_alloc;
+  }
+
+  q->buf.arr = arr;
+  q->buf.len = cap;
+
+  q->tail = 0;
+  q->head = 0;
+  q->empty = true;
+
+  pthread_mutex_init(&q->mutex, NULL);
+  pthread_cond_init(&q->cond, NULL);
+
+  *out = q;
+
+  return lnm_err_ok;
+}
+
+void lnm_loop_queue_push(lnm_loop_queue *q, lnm_loop_conn *conn) {
+  pthread_mutex_lock(&q->mutex);
+
+  while (q->head == q->tail && !q->empty) {
+    pthread_cond_wait(&q->cond, &q->mutex);
+  }
+
+  q->buf.arr[q->head] = conn;
+
+  // Make sure the index wraps around
+  q->head = (q->head + 1) % q->buf.len;
+  q->empty = false;
+
+  // Unlock mutex and signal to waiting threads
+  pthread_mutex_unlock(&q->mutex);
+  pthread_cond_signal(&q->cond);
+}
+
+lnm_loop_conn *lnm_loop_queue_pop(lnm_loop_queue *q) {
+  pthread_mutex_lock(&q->mutex);
+
+  while (q->empty) {
+    pthread_cond_wait(&q->cond, &q->mutex);
+  }
+
+  lnm_loop_conn *out = q->buf.arr[q->tail];
+
+  q->tail = (q->tail + 1) % q->buf.len;
+  q->empty = q->tail == q->head;
+
+  // Unlock mutex and signal to waiting threads
+  pthread_mutex_unlock(&q->mutex);
+  pthread_cond_signal(&q->cond);
+
+  return out;
+}
+
+void lnm_loop_worker_run(void *arg) {
+  lnm_loop *l = arg;
+  lnm_loop_queue *q = l->wq;
+
+  // Get thread ID by incrementing counter
+  pthread_mutex_lock(&l->threads.mutex);
+
+  int thread_id = l->threads.worker_count;
+  l->threads.worker_count++;
+
+  pthread_mutex_unlock(&l->threads.mutex);
+
+  while (1) {
+    lnm_loop_conn *conn = lnm_loop_queue_pop(q);
+    lnm_ldebug("loop", "worker %i processing fd %i", thread_id, conn->fd);
+
+    lnm_loop_conn_advance(l, conn);
+    lnm_loop_conn_schedule(l, conn);
+  }
+}